METHOD AND SYSTEM FOR CONFIDENTIAL STRING-MATCHING AND DEEP PACKET INSPECTION

Abstract

Provided is a system and method for confidential string-matching and confidential deep-packet inspection. The method includes: receiving encrypted ciphertexts from a first computing device; windowing a text corpus and applying a hash; performing binning and splitting on the corpus set of hashes; performing batching on the binned and split corpus set of hashes; determining match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes; and communicating the match ciphertexts to the first computing device, the confidential string matching determinable by the first computing device by: decrypting the match ciphertexts, determining from the decryption output, if the hash value for each pattern window matches the hash value for any corpus windows and if the matched windows are adjacent in the corpus.

Description

TECHNICAL FIELD

The following relates generally to data processing, and more specifically, to a method and system for confidential string-matching and deep packet inspection.

BACKGROUND

String-matching is a common data processing problem of searching for a pattern of characters within a body of text. The general string-matching approaches do not generally apply any security or privacy features. Confidential string-matching is a variation of the problem where the pattern and the result must remain confidential in an execution environment. In confidential string-matching, the character pattern is encrypted and sent over to a host who performs the confidential string-matching on a body of text. The result of such matching remains confidential, and in many cases, the host does not learn whether there was a match or not. Only the party which has the private key can decrypt and see the match results. This is a significantly harder technical challenge than the general string-matching problem.

SUMMARY

In an aspect there is provided a method for confidential string-matching of input patterns to a corpus, each input pattern comprises elements from an alphabet, the corpus comprises elements from the alphabet, the method executed on at least one processing unit of a second computing device, the method comprising: receiving encrypted ciphertexts from a first computing device; windowing the corpus and applying a hash function on the corpus windows to form a corpus set of hashes; performing binning on the corpus set of hashes; performing batching on the binned corpus set of hashes; determining match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes; and communicating the match ciphertexts to the first computing device to determine the confidential string matching.

In a particular case of the method, the first computing device determines the confidential string matching by: decrypting the match ciphertexts, and determining, from the output of the decryption, if the hash value for each pattern window matches the hash value for any corpus windows and if the matched windows are adjacent in the corpus.

In another case of the method, the first computing device having determined the encrypted ciphertexts by: determining a hashed set from windowing the input patterns and determining hash values for each pattern window; performing binning by placing items in the hashed set into a table; batching the elements of the table into polynomials; and performing encryption on the polynomials to generate the encrypted ciphertexts;

In yet another case of the method, performing the binning comprises mapping all elements in the corpus set of hashes to one of a plurality of bins, and splitting the elements in each of the bins into one of a plurality of sub-bins.

In yet another case of the method, performing the batching comprises determining coefficients of a polynomial for each sub-bin, and batching the coefficients of each polynomial into a plaintext polynomial.

In yet another case of the method, evaluating the homomorphic encryption circuit comprises homomorphically evaluating a function comprising a combination of the ciphertexts and the plaintext polynomials associated with each of the sub-bins.

In yet another case of the method, the method further comprising performing compression on the evaluated function.

In yet another case of the method, the homomorphic encryption comprises an indistinguishability under chosen-plaintext attack (IND-CPA) secure homomorphic encryption scheme, and wherein the parameters for the IND-CPA secure homomorphic encryption scheme are communicated between the second computing device and the first computing device.

In yet another case of the method, evaluating the homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes further determines a label ciphertext that encodes a set of these locations of the match ciphertexts.

In another aspect, there is provided a system for confidential string-matching of input patterns to a corpus, each input pattern comprises elements from an alphabet, the corpus comprises elements from the alphabet, the system comprises at least one processing unit and a data storage media in communication with the at least one processing unit, the at least one processing unit and the data storage media part of a second computing device, the at least one processing unit configured to execute: a second input module to receive encrypted ciphertexts from a first computing device; a second pre-processing module to window the corpus and applying a hash function on the corpus windows to form a corpus set of hashes, to perform binning on the corpus set of hashes, and to perform batching on the binned corpus set of hashes; a homomorphic evaluation module to determine match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes; and a network interface to communicate the match ciphertexts to the first computing device to determine the confidential string matching.

In a particular case of the system, the first computing device comprises at least one processing unit, and wherein determining the confidential string matching is performed by executing on the at least one processing unit: decrypting the match ciphertexts, and determining, from the output of the decryption, if the hash value for each pattern window matches the hash value for any corpus windows and if the matched windows are adjacent in the corpus.

In another case of the system, the first computing device having determined the encrypted ciphertexts by executing on the at least one processing unit: determining a hashed set from windowing the input patterns and determining hash values for each pattern window; performing binning by placing items in the hashed set into a table; batching the elements of the table into polynomials; and performing encryption on the polynomials to generate the encrypted ciphertexts;

In yet another case of the system, performing the binning comprises mapping all elements in the corpus set of hashes to one of a plurality of bins, and splitting the elements in each of the bins into one of a plurality of sub-bins.

In yet another case of the system, performing the batching comprises determining coefficients of a polynomial for each sub-bin, and batching the coefficients of each polynomial into a plaintext polynomial.

In yet another case of the system, evaluating the homomorphic encryption circuit comprises homomorphically evaluating a function comprising a combination of the ciphertexts and the plaintext polynomials associated with each of the sub-bins.

In yet another case of the system, the homomorphic evaluation module further performs compression on the evaluated function.

In yet another case of the system, the homomorphic encryption comprises an indistinguishability under chosen-plaintext attack (IND-CPA) secure homomorphic encryption scheme, and wherein the parameters for the IND-CPA secure homomorphic encryption scheme are communicated between the second computing device and the first computing device.

In yet another case of the system, evaluating the homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes further determines a label ciphertext that encodes a set of these locations of the match ciphertexts.

In another aspect, there is provided a method for confidential deep-packet inspection of input signatures to traffic over a network, the input signatures comprise elements from an alphabet, the method executed on at least one processing unit of a second computing device, the method comprising: second input module receiving encrypted ciphertexts from a first computing device, the first computing device having determined the encrypted ciphertexts by: determining a hashed set from windowing the input signatures and determining hash values for each signature window, performing binning by placing items in the hashed set into a table, batching the elements of the table into polynomials, and performing encryption on the polynomials to generate the encrypted ciphertexts; buffering the network traffic in a specified time interval to form a list of packets, each packet comprising a packet body and a packet 5-tuple, the packet body and the packet 5-tuple comprise elements from the alphabet; windowing each packet body and applying a hash function on the packet windows to form a packet set of hashes; performing binning and splitting on the packet set of hashes; performing batching on the binned and split packet set of hashes; determining match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched packet set of hashes; and communicating the match ciphertexts and the packet 5-tuples to the first computing device, the confidential deep-packet inspection determinable by the first computing device by: decrypting the match ciphertexts, determining, from the output of the decryption, if the hash value for each signature window matches the hash value for any packet windows, determining if the matched windows are adjacent in a packet, and determining the associated 5-tuple from the received packet 5-tuples.

In another aspect, there is provided a system for confidential deep-packet inspection of input signatures to traffic over a network, the input signatures comprise elements from an alphabet, the system comprises at least one processing unit and a data storage media in communication with the at least one processing unit, the at least one processing unit and the data storage media part of a second computing device, the at least one processing unit configured to execute: a second input module to receive encrypted ciphertexts from a first computing device, the first computing device having determined the encrypted ciphertexts by determining a hashed set from windowing the input signatures and determining hash values for each signature window, performing binning by placing items in the hashed set into a table, batching the elements of the table into polynomials, and performing encryption on the polynomials to generate the encrypted ciphertexts; a packet module to buffer the network traffic in a specified time interval to form a list of packets, each packet comprising a packet body and a packet 5-tuple, the packet body and the packet 5-tuple comprise elements from the alphabet, and to window each packet body and applying a hash function on the packet windows to form a packet set of hashes; a second pre-processing module to perform binning and splitting on the packet set of hashes, and to perform batching on the binned and split packet set of hashes; a homomorphic evaluation module to determine match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched packet set of hashes; and a network interface to communicate the match ciphertexts and the packet 5-tuples to the first computing device, the confidential deep-packet inspection determinable by the first computing device by: decrypting the match ciphertexts, determining, from the output of the decryption, if the hash value for each signature window matches the hash value for any packet windows, determining if the matched windows are adjacent in a packet, and determining the associated 5-tuple from the received packet 5-tuples.

These and other embodiments are contemplated and described herein. It will be appreciated that the foregoing summary sets out representative aspects of systems and methods to assist skilled readers in understanding the following detailed description.

BRIEF DESCRIPTION OF THE DRAWINGS

The features of the invention will become more apparent in the following detailed description in which reference is made to the appended drawings wherein:

FIG. 1 is a schematic diagram of a system for confidential string-matching, in accordance with an embodiment;

FIG. 2 is a schematic diagram showing an example embodiment of a first computing device of the system of FIG. 1;

FIG. 3 is a schematic diagram showing an example embodiment of a second computing device of the system of FIG. 1;

FIG. 4 is a diagrammatic example of binning and batching in labelled PSI protocol;

FIG. 5 is a diagrammatic example of splitting in labelled PSI protocol;

FIG. 6 illustrates a diagram of an example high-level overview of an approach to confidential string-matching by the system of FIG. 1;

FIGS. 7A and 7B are a flowchart of a method for confidential string-matching, in accordance with an embodiment;

FIG. 8 illustrates a diagram of an example high-level overview of an approach to confidential deep packet inspection by the system of FIG. 1;

FIGS. 9A and 9B are a flowchart of a method for confidential deep packet inspection using confidential string-matching, in accordance with an embodiment; and

FIG. 10 is a diagram showing an example of confidential string-matching with R-extension.

DETAILED DESCRIPTION

Embodiments will now be described with reference to the figures. For simplicity and clarity of illustration, where considered appropriate, reference numerals may be repeated among the Figures to indicate corresponding or analogous elements. In addition, numerous specific details are set forth in order to provide a thorough understanding of the embodiments described herein. However, it will be understood by those of ordinary skill in the art that the embodiments described herein may be practiced without these specific details. In other instances, well-known methods, procedures and components have not been described in detail so as not to obscure the embodiments described herein. Also, the description is not to be considered as limiting the scope of the embodiments described herein.

Various terms used throughout the present description may be read and understood as follows, unless the context indicates otherwise: “or” as used throughout is inclusive, as though written “and/or”; singular articles and pronouns as used throughout include their plural forms, and vice versa; similarly, gendered pronouns include their counterpart pronouns so that pronouns should not be understood as limiting anything described herein to use, implementation, performance, etc. by a single gender; “exemplary” should be understood as “illustrative” or “exemplifying” and not necessarily as “preferred” over other embodiments. Further definitions for terms may be set out herein; these may apply to prior and subsequent instances of those terms, as will be understood from a reading of the present description.

Any module, unit, component, server, computer, terminal, engine or device exemplified herein that executes instructions may include or otherwise have access to computer readable media such as storage media, computer storage media, or data storage devices (removable and/or non-removable) such as, for example, magnetic disks, optical disks, or tape. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, program modules, or other data. Examples of computer storage media include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by an application, module, or both. Any such computer storage media may be part of the device or accessible or connectable thereto. Further, unless the context clearly indicates otherwise, any processor or controller set out herein may be implemented as a singular processor, as a plurality of processors, as a multi-core and/or multi-threaded processors, or the like. The plurality of processors may be arrayed or distributed, and any processing function referred to herein may be carried out by one or by a plurality of processors, even though a single processor may be exemplified. Any method, application or module herein described may be implemented using computer readable/executable instructions that may be stored or otherwise held by such computer readable media and executed by the one or more processors.

The following relates generally to data processing, and more specifically, to a method and system for confidential string-matching and deep packet inspection.

The string-matching problem of searching for instances of a pattern within a body of text is a problem with multiple deterministic and approximate approaches for both the single-pattern and multiple-pattern variants of the problem. In an example, the string matching operation is an important operation in the field of bioinformatics.

A naive approach consists of performing a character-wise string comparison between an input pattern and every possible pattern-sized substring in the body of text by creating a window with a size equal to the pattern size and sliding this window over the text one character at a time. In a variation of this approach, a fingerprint for the pattern can be generated and for each of the possible pattern-sized substrings in the body. A fingerprint is an identifier that attempts to uniquely identify a string but is smaller in size than the string itself. The pattern fingerprint is compared to all other fingerprints and a character-wise comparison is performed only if the fingerprints match. The efficiency of this approach depends on two ideas:

• • 1) The fingerprint comparison being faster than character-wise string comparison; and
  • 2) The fingerprinting approach producing highly unique fingerprints to reduce character-wise string comparisons between strings that do not match.

Generalized versions of the fingerprinting approach allow for multiple-pattern searches by fingerprinting fixed-size substrings for all patterns and text.

A particular approach to single-pattern matching exploits the idea that while sliding the pattern on the text, instead of sliding the pattern window by one character like in the naive solution, multi-character slides can be made based on knowledge from comparing just the tail character from the text in the window with characters in the pattern. This allows for sub-linear performance with respect to the text length.

A particular approach to multiple-pattern searches involves the use of a finite-state machine that allows for simultaneous matching of all patterns; thus, reducing the performance impact stemming from searching for multiple patterns. The state machine is constructed using characters from all patterns with links representing the order of characters in the patterns. Supplementary links can be used for fast transition between failed string matches during traversing.

A substantial problem with the above approaches is that progression during execution is reliant upon knowing the partial results of comparisons between pattern data and text data. Therefore, the above approaches cannot be trivially transformed into a form where the data of the pattern or the result remains confidential during execution; thus, failing to satisfy general constraints of confidential string-matching.

A Private Set Intersection (PSI) problem is a problem where two parties, the Receiver and the Sender, each own a private set of text and the Receiver only learns an intersection between the two sets while the Sender learns nothing.

One approach to the PSI problem involves the use of Oblivious Pseudo Random Functions (OPRF), where the Sender holds the key for OPRF. In this approach, the Receiver applies OPRF, by performing the OPRF protocol with Sender, on its set Y to form a new set Y′. The Sender applies OPRF to its own set X locally and sends the new set X′ to the Receiver. The Receiver can now find the intersection of the new sets X′ and Y′, therefore learning the intersection of the set X and Y. The Sender learns nothing. Due to requiring sending of the sets, this approach is inefficient communication-wise; especially when the sets are unbalanced in size.

Another approach to the PSI problem involves using Homomorphic Encryption (HE) to solve an unbalanced PSI problem more efficiently. It transforms the element comparisons into arithmetic circuits that can be evaluated with HE operations. It also has optimizations for efficient mapping and comparison of the sets in HE-compatible structures.

While the above approaches to the PSI problem effectively allow for searching contents of a confidential set in another set, while keeping the result confidential, these approaches accept as inputs databases with well defined fields. Thus, these approaches cannot be applied directly to perform string-matching where one searches for one or more occurrences of a string in a larger string. Moreover, the multiple-pattern variant of string-matching complicates these approaches even further.

For deep packet inspection (DPI), network intrusion systems are deployed as middleboxes on networks that inspect the traffic and enforce a set of rules based on the pattern data in the traffic. These rules consist of a set of actions to perform based on whether a set of signatures exist in the network traffic. Encrypted Network Traffic Searching is the problem of executing searches for signatures on network traffic that utilize communication protocols that encrypt the traffic content for privacy and security.

Some approaches for DPI involve middleboxes that function by decrypting the encrypted traffic before performing the search for signatures. Other approaches utilize purpose-built encryption schemes to allow for searching while keeping the traffic and the signatures confidential. Some other approaches even allow for the actions within the rules to remain confidential until the middlebox needs to perform them. The issue with these approaches is that, while they provide an effective form of matching signatures that are encrypted, they require the network traffic to be encrypted as well. More importantly, instead of keeping the results of the search confidential, they expose the results in order to compute or perform actions from the respective rule; hence, failing the requirements for confidential string-matching.

In the following disclosure, the following notation will be referred to:

• • For string-matching specifically: P is a pattern set known by the Receiver; s is a text corpus known by the Sender.
  • For DPI specifically: Z is a signature set known by an Agency (Receiver in a Secure Environment); D is a packet set known by the Host (Sender in an Insecure Environment) containing ordered pairs of (a,b) where a is a packet body and b is a corresponding packet 5-tuple.
  • γ is a minimum possible length for s and for items in P, Z, D_a.
  • ω is a size of a rolling hash window to apply to s and items in P, Z, D_a where ω≤γ.
  • X is the Sender/Host's set of fingerprints while Y is the Receiver/Agency's set. It is assumed that N_X»N_Y.
  • σ is a bit length of items in X and Y.
  • l represents a label of an element in X.
  • n is a ring dimension in HE (a power of 2).
  • q is a ciphertext modulus.
  • t is a plaintext modulus.
  • F_t is a field for a plaintext polynomial.
  • κ denotes a computational security parameter.
  • λ denotes a statistical security parameter.
  • m is a number of bins in a hash table.
  • α is a number of partitions that a Host/Sender set X is split into.
  • ψ is a maximum number of elements in a bin of a hash table, after α-splitting.

A particular approach to string-matching is called the Rabin-Karp algorithm, which takes advantage of fingerprinting to reduce the amount of character-level comparisons involved in a naive string-matching approach. For a pattern string p and a text string s, a determination of a fingerprint for p and the fingerprints for each p-sized substring in s are made. The fingerprint of p is compared to each fingerprint from s, only comparing characters between the strings when their respective fingerprints match. This approach generally performs better than the naive approach because of the selection of a fingerprinting algorithm that can provide, for example, the following properties:

• • 1) Virtual Uniqueness: The probability of generating the same fingerprint for two different strings should ideally be negligible. The higher this probability of collision is, the more character-level comparisons need to be performed due to the possibility of false positives.
  • 2) Efficient Computation: The fingerprint calculation and comparison for string of size m should be considerably more efficient than the character-by-character comparison of two strings of size m (not represented as fingerprints).
  • 3) Computation Reuse: The calculation of the fingerprint for substring ζ_i in s should be significantly accelerated if provided with the fingerprint for substring ζ_i-1 in s.

Generally, fingerprinting algorithms, such as hash functions, only satisfy properties 1 and 2 with some cryptographic variants failing to even satisfy property 2. In contrast, a rolling hash, such as the Rabin fingerprint, is generally able to satisfy all three properties adequately. The Rabin fingerprint treats the input string as a polynomial of bits (a polynomial over Galois field GF(2)) and generates the hash by dividing the input by a pre-selected irreducible polynomial (over GF(2)). The ability to update the Rabin fingerprint using only the entry and exit bytes allows it to satisfy property 3. This leads to higher efficiency for large pattern lengths since the complexity of the fingerprints for the large patterns is effectively the same as complexity for the short pattern fingerprint calculations; i.e., the complexity of the fingerprint determination is constant with respect to pattern length.

Additionally, character-level string comparison may also be excluded if the fingerprinting algorithm satisfies virtual uniqueness and the level of false positives received is tolerable for an intended application. The exclusion of character-level string comparison also allows for a parallel implementation in hardware as the algorithm never needs to revisit the data in s if it is being streamed in.

For single-pattern string-matching, the Rabin-Karp algorithm is usually slower than other algorithms, such as Knuth-Morris-Pratt, as these algorithms gain more information about s for every mismatch and adapt accordingly. Rabin-Karp also suffers from having a worst case behaviour equivalent to the naive approach if the fingerprinting algorithm has high collision probability. However, for multi-pattern string-matching, the Rabin-Karp algorithm may be more appropriate because it compares fingerprints for many patterns. It may additionally compress the pattern fingerprints into a Bloom Filter for constant-time approximate fingerprint comparison at the cost of higher false-positive probability. To adapt for differing lengths of patterns, fingerprints of substrings in s need to be generated for each pattern length. Alternatively, a more generalized version of the algorithm can be used which makes use of rolling windows. In this version, fingerprints of window-sized substrings of s and each p∈P, are generated and compared (where P is the pattern set and s is the text string). The window size is selected to be smaller than or equal to size of the shortest p.

Labelled private set intersection (PSI) is a protocol that allows two parties, Sender and Receiver, to perform private set intersection on their respective sets. The Receiver is assumed to hold a set (in an example, a list of customer names). The Sender is assumed to hold another set, where each element in the set is associated with a label. This label can represent any construct, in this example, ranging from a data field associated with each customer name to a handle for objects (such as documents, images, or the like). The Labelled PSI protocol returns to the Receiver, the intersection of the Sender's and Receiver's set, and the corresponding labels of the intersection. This protocol leverages HE, such that the computation and communication will generally reveal no information about the result to the Sender.

To perform the PSI protocol, the intersection of two sets is performed by evaluating a “Match Function”:

F(x)=(x*(x−y₁)*(x−y₂)* . . . *(x−y_NY)

where x is an element in the Receiver's Set, and y₁, y₂, . . . , y_NY are all elements in the Sender's set. It can be seen that when x exists in the Sender's set, F(x) will output zero, otherwise it is an arbitrary value.

The corresponding labels for the intersection are obtained by evaluating a “Label Function”:

G(x)={l if x exists in set; arbitrary_value otherwise}

G(x) is created by interpolating a polynomial on all pairs of (y_i, l_i), where y_i is an element in the Sender's set and l_i is its associated label. So when x exists in the Sender's set, G(x) will output the corresponding label; otherwise it will output an arbitrary value.

The encrypted results of F(x) and G(x) are sent back to the Receiver for decryption. When F(x) is zero, the element exists in the set, and that G(x) is the correct label associated with that element; otherwise the value in G(x) is an arbitrary value. The above procedure is repeated for each element in the Receiver's set to find the entire intersection and associated labels. The above procedure can be directly mapped to HE by encoding an element x into a ciphertext, and mathematical operations in F(x) and G(x) are performed using corresponding HE mathematical operations.

In some cases, the Labelled PSI procedure can use a variety of techniques to reduce the communication and computation costs, including binning, batching, windowing, and splitting.

For batching, each query element is encoded in a single ciphertext. However, depending on the choice of HE security parameters, each ciphertext can hold many coefficients based on the degree of the polynomial encoded. The mathematical operations on the HE ciphertext are performed coefficient-wise, therefore the evaluation of Match Function and Label Function can be applied to all coefficients of a ciphertext simultaneously. Batching the query elements reduces the communication cost and memory footprint significantly.

Binning is a technique to reduce the search space for a query in Labelled PSI. In a basic case, each Match Function and Label Function evaluation is effectively searching an element in the entire Sender's set. To reduce the search space, the Sender and Receiver agree upon a set of hash functions, then the Sender places its elements into a set of bins by hashing each element. Similarly, the Receiver will use the same set of hash functions to perform hashing on its set to form its own set of bins. In some cases, cuckoo-hashing can be used instead of regular hashing to distribute across the hash table more densely so that the space in the query ciphertext can be used more effectively. The effect of binning is that an element in the Receiver's bin can only exist in the corresponding Sender's bin, so a search need only be conducted in corresponding bins instead of the entire set. Binning can be combined with batching to perform multiple Match and Label evaluations with a single query ciphertext as long as the Receiver's database elements are placed into different bins (with no collisions), as illustrated in the example of FIG. 4.

The Splitting technique is used to break up each bin into a smaller bins, and a query is performed on each of the smaller bins. The results are collected from the different sub-bins, as illustrated in the example of FIG. 5. The effect of splitting is three-fold:

• • Query on each smaller bin is independent from one another, thus can be parallelized.
  • Splitting helps reducing the bin load of each bin, which can result in a smaller-sized interpolation to create G(x).
  • Smaller bin load leads to less noise growth when evaluating F(x) and G(x).

However, the splitting technique will generally increase the number of result ciphertexts by the same factor α.

The match function F(x) can be expanded as follows:

F(x)=(x−y₁)*(x−y₂)* . . . *(x−y_ψ)=a₁+a₂x+a₃x²+ . . . +_ψ+1x^ψ

The a₁, a₂, a₃, . . . , a_ψ+1 are coefficients of the powers of x after the expansion, which depend only on individual bins from the Sender's set. This minimizes the number of consecutive multiplications required in ciphertext given that some or all the powers of x (i.e. x, x², x³, . . . , x^ψ) can be pre-computed in plaintext and encrypted individually at the Receiver before sending to the Sender.

The Receiver sends the query ciphertext x to the Sender. To avoid the Sender needing to exponentiate the query ciphertext, the Receiver computes the powers of the query in plaintext and encrypts them before sending it to the Sender. However, sending all powers of x can become prohibitively large for communication. The Receiver can choose to send windows of x instead, e.g. x, x², x⁴, x⁸, . . . . The Sender can then compute the missing powers.

In most cases, HE operations, such as addition and multiplication, incur noise growth on the ciphertexts. Excessive noise growth can result in decryption failure. The splitting and binning operations can be used to decrease the degree of F(x) and G(x) polynomial functions.

In some cases, some parts of the protocol can be pre-computed to speed up subsequent processing, for example online processing. Among these parts are:

• • Coefficients of Match Function after expansion F(x): a₁, a₂, a₃, . . . , a_ψ+1
  • Coefficient of Interpolation Polynomial G(x): k₁, k₂, k₃, . . . , k_ψ+1
  • Powers of query ciphertext x: x, x², x³, . . . , x^ψ

In the present embodiments, any suitable homomorphic encryption (HE) can be used; for example, Fully Homomorphic Encryption (FHE), Somewhat Homomorphic Encryption (SHE), Levelled Homomorphic Encryption (LHE), Partial Homomorphic Encryption (PHE), and the like.

Advantageously, embodiments of the present disclosure provide confidential string-matching protocol using HE; such as integrating the Rabin-Karp algorithm and Labelled PSI protocol to form the confidential string-matching protocol. Additionally, embodiments of the present disclosure provide confidential deep packet inspection using HE by applying multiple optimization techniques such that the confidential string-matching protocol can be used in deep packet inspection over network traffic. Additionally, embodiments of the present disclosure provide compression techniques to reduce the size of the encrypted results, which results in a reduced communication overhead. Additionally, embodiments of the present disclosure provide multiple parallelization and pipelining optimizations to enable use in practical applications.

In embodiments of the present disclosure, the Rabin Karp algorithm and the Labelled PSI are integrated by fingerprinting the substrings of the patterns and the text corpus, encoding the location of each fingerprint in the text corpus as a label, and performing Labelled PSI on the two fingerprint sets to obtain the matched fingerprints and their locations in the original text string. These results can be post-processed to know which pattern matched where in the text. This integration advantageously enables confidential string-matching using HE.

The string-matching protocol of embodiments described herein can be utilized in inspecting network packets in an insecure environment. Embodiments of the present disclosure also provide optimization techniques to provide an approach that can be run over high speed networks without leaking, (1) the set of signatures being searched for and (2) whether there was a match or not.

The Labelled PSI protocol can contain an α-splitting optimization to reduce the search space; however, generally at a cost of a times larger result size. As the results are all encrypted, the amount of ciphertexts the protocol generates may become prohibitive in a practical setting. Embodiments of the present disclosure provide a technique to compress the number of result ciphertexts to achieve higher efficiency in real applications.

Embodiments of the present disclosure exploit the fact that the a pairs of ciphertext results from the splitting approach generally contain the same information as the single pair of ciphertext results obtained without splitting. Therefore, they are compressible from an information theory perspective. This fact can be used to construct an arithmetic circuit that operates on the original result ciphertexts to reduce the results to a constant number of ciphertexts with respect to the length of text string being searched.

Embodiments of the present disclosure leverage parallelization, multi-threading, GPU computing, and pipelining techniques to enable practical, high-performance confidential string-matching. These techniques can be used to accelerate and schedule multiple sections of the approach, for example, without the loss of generality, polynomial interpolation, and HE operations.

Referring now to FIG. 1, a system 100 for confidential string-matching, in accordance with an embodiment, is shown. In this embodiment, the system 100 includes a first computing device 102 (in the confidential string matching embodiment, referred to as a Receiver) and a second computing device 152 (in the confidential string matching embodiment, referred to as a Sender) communicating over a network 130 (for example, a local area network or the Internet). The computing devices can be any suitable device; for example, a desktop computer, a laptop computer, a smartphone, a tablet computer, a mobile device, a smartwatch, or the like. In further embodiments, aspects or functions of the system can be run on further computing devices, such as a separate server. In some embodiments, the components of the Receiver computing device 102 and the Sender computing device 152 are each stored by and executed on a single computer system. In other embodiments, the components of the Receiver computing device 102 and the Sender computing device 152 are each distributed among two or more computer systems that may be locally or remotely distributed.

FIG. 2 shows an example embodiment of the first computing device 102 including various physical and logical components. As shown, the first computing device 102 has a number of physical and logical components, including a central processing unit (“CPU”) 110 (comprising one or more processors), random access memory (“RAM”) 104, an input interface 106, an output interface 108, a network interface 114, non-volatile storage 112, and a local bus 114 enabling CPU 110 to communicate with the other components. CPU 110 executes an operating system, and various modules, as described below in greater detail. RAM 104 provides relatively responsive volatile storage to CPU 110. The input interface 106 enables an administrator or user to provide input via an input device, for example a keyboard and mouse. The output interface 108 outputs information to output devices, such as a display and/or speakers. The network interface 114 permits communication with other systems, such as the Sender 152 over the network 130, or other computing devices and servers remotely located from the system 100, such as for a typical cloud-based access model. Non-volatile storage 112 stores the operating system and programs, including computer-executable instructions for implementing the operating system and modules, as well as any data used by these services. Additional stored data, as described below, can be stored in a database 116. During operation of the system 100, the operating system, the modules, and the related data may be retrieved from the non-volatile storage 112 and placed in RAM 104 to facilitate execution. In an embodiment, the first computing device 102 further includes a number of conceptual modules, including a first input module 118, a first pre-processing module 120, and a post-processing module 122. In some cases, the modules 118, 120, 122 can be executed on the CPU 110. In further cases, some of the functions of the modules 118, 120, 122 can be executed on a server, on cloud computing resources, or other devices. In some cases, some or all of the functions of any of the modules 118, 120, 122 can be run on other modules.

FIG. 3 shows an example embodiment of the second computing device 152 including various physical and logical components. As shown, similar to the first computing device 102, the second computing device 152 has a number of physical and logical components, including a central processing unit (“CPU”) 160 (comprising one or more processors), random access memory (“RAM”) 154, an input interface 156, an output interface 158, a network interface 154, non-volatile storage 162, and a local bus 164 enabling CPU 160 to communicate with the other components. CPU 160 executes an operating system, and various modules, as described below in greater detail. RAM 154 provides relatively responsive volatile storage to CPU 160. The input interface 156 enables an administrator or user to provide input via an input device, for example a keyboard and mouse. The output interface 158 outputs information to output devices, such as a display and/or speakers. The network interface 164 permits communication with other systems, such as the second computing device 152 over the network 130, or other computing devices and servers remotely located from the system 100, such as for a typical cloud-based access model. Non-volatile storage 162 stores the operating system and programs, including computer-executable instructions for implementing the operating system and modules, as well as any data used by these services. Additional stored data, as described below, can be stored in a database 166. During operation of the system 100, the operating system, the modules, and the related data may be retrieved from the non-volatile storage 162 and placed in RAM 154 to facilitate execution. In an embodiment, the Receiver 102 further includes a number of conceptual modules, including a second input module 168, a second pre-processing module 170, a homomorphic evaluation module 172, and a packet module 174. In some cases, the modules 168, 170, 172, 174 can be executed on the CPU 160. In further cases, some of the functions of the modules 168, 170, 172, 174 can be executed on a server, on cloud computing resources, or other devices. In some cases, some or all of the functions of any of the modules 168, 170, 172, 174 can be run on other modules.

In further cases, the CPU 110, 160 can be a combination of multi-core CPU and GPU to achieve high performance. Alternatively, the CPU 110, 160 can be implemented on custom ASICs and/or FPGAs.

In further cases, the network 130 data stream can be statistically de-multiplexed into multiple streams and fed to identical hardware units. This can lead to higher processing capacity and therefore ability to handle higher input bandwidth.

The system 100 is able to perform confidential string-matching as a variation of the string-matching problem in which a character pattern is encrypted and sent over to a Sender that performs the confidential string-matching on a body of text. It is a general requirement that the result of such matching must remain confidential and the Sender cannot learn whether there was a match or not. Only the party which has the private key can decrypt and see the match results. In some examples, confidential string-matching is particularly useful in the following scenarios:

• • A Digital Safety Deposit Box: confidential string-matching can be used by a digital container that can store sensitive documents (e.g., Passports, Security Certificates, Land Deeds, etc.) in an encrypted or unencrypted format, such that the owner can perform a search with encrypted queries on the contents of the documents (or indexes or metadata) without revealing the queries or the results.
  • Confidential web search: confidential string-matching can be used by a web search engine that can perform searches on website indexes without the website host being aware of the query string or the results, protecting the privacy of the user.
  • Confidential access to a database: confidential string-matching can be used to permit access to a database that is public, while hiding the query and result.
  • Confidential access to an encrypted database: confidential string-matching can be used to securely access an encrypted database. whereas an example, the United States Security and Exchange Commission (SEC) authority may wish to do secure insider trading audits on metadata associated with encrypted commodity trading data.
  • IP theft detection: confidential string-matching can be used by an agency to perform a search for unpublished or protected Intellectual Property they own to check for leaks or infringements without revealing information to the search execution space.
  • Confidential Search Engine: confidential string-matching can be used to perform encrypted queries on encrypted or plaintext data and the result remains encrypted and can only be decrypted by the owner of the private key.
  • Confidential R&D or Legal Document Search: confidential string-matching can be used to securely search patent metadata and abstracts protecting the privacy of company R&D teams and legal teams for the purpose of not revealing the specific technical subject matter of interest or under development by the company.
  • Confidential Geolocation Quarantine Compliance: In an epidemic, confidential string-matching can be used by a user who is to be quarantined at home. Where their cell phone GPS coordinates can be encrypted and used as a pattern in confidential string matching to verify against their home, doctor's office, hospital, or pharmacist geolocations to securely verify that they are fulfilling their quarantine obligations.
  • Contact Tracing: Confidential string matching can be used to find the intersection of GPS coordinates (or equivalent location specifier). This can be useful for finding people and/or objects (e.g., autos, trucks, shipping containers, and the like) that were in close proximity to a person and/or location of interest (e.g., a person who has infectious disease).

For string-matching to be secure and confidential, the confidentiality of the patterns and the match results generally need to remain secure and confidential during runtime. In order to achieve this utilizing Homomorphic Encryption (HE), the confidential string-matching algorithm can be represented using arithmetic circuits. For a confidential pattern p and non-confidential (or confidential) text s, an expression that can be used is:

f(p)=(p₁−s_i)*(p₂−s_(i+1))*(p₃−s_(i+2))* . . . *(p_Np−s_(i+Np))

where p_i and s_i represents the i^th character in a string p and s, respectively. This expression is to be determined for each i, where 1≤i≤N_s−N_p+1; which means the result size is proportional to the length of s. The length of p is non-confidential. The circuit depth is proportional to the length of p. Without even considering multi-pattern searches, the HE operation performance and ciphertext size of other approaches quickly make the above expression impractical as the length of p or s is increased. Advantageously, embodiments of the present disclosure provide a more efficient approach in order to be used in practical applications.

The matching component of string-matching algorithms, such as Knuth-Morris-Pratt (KM) and Boyer-Moore (BM), is defined by the representation and traversal of complex structures; such as deterministic finite automatons (DFA). Representation and traversal of these structures as an arithmetic circuit, while ensuring the pattern remains confidential, is difficult and non-trivial. Meanwhile, the matching component of the Rabin-Karp (RK) algorithm, described herein, includes an approach that can be used to integrate with HE. If the fingerprint algorithm in RK can offer negligible collision probability, the character-level search could effectively be discarded. The matching component could then be described as a search for a fingerprint (of p) in a set of fingerprints (from s). Enforcing the requirement of keeping p confidential turns this into the Private Set Intersection (PSI) problem where at least one set is required to be private (in some cases, both sets can get encrypted if needed). For the case of unbalanced PSI, where one set is much smaller than the other, the system 100 can advantageously integrate RK with the Labelled PSI to implement confidential string-matching.

FIG. 6 illustrates a diagram of an example high-level overview of an approach to confidential string-matching by the system 100. In this approach, in addition to knowing whether p exists in s, the system 100 can also identify the locations in s where p matched. The label in the labelled PSI algorithm is used to encode the set of these locations as a label. Additionally, since the system 100 addresses the multi-pattern confidential string-matching problem, a generalized version of Rabin-Karp can be used. In a particular case, this involves rolling a fixed-size window over each pattern p in P and generating a fingerprint for each unique window to form the confidential set of fingerprints, Y. The same rolling window process can be applied to the body of text s, to form the non-confidential set of fingerprints X. This process effectively translates to searching for windows from each p instead of searching for p itself. To account for this, in most cases, a post-decryption step can be performed. In this step, if all windows for a p matched, then the locations of these windows are compared to check if the matched windows were consecutively adjacent in s; hence, confirming whether p exists in s. If the locations in s are not a requirement of the problem and some level of false positives are tolerable in the intended application, then the label encoding and offset comparison can be excluded in order to save on computation and communication at the cost of potential false positives.

In some embodiments, to determine input transformations, for each pattern p∈P, a fixed-size window ω is rolled over p, giving (len(p)−len(w)+1) fingerprints for each p (where the function len(a) returns the length of the string a). Each fingerprint can be treated as an element in the Receiver's set Y; and thus forms the Receiver's set Y. Additionally, a fixed-size window ω can be rolled over text s, giving (len(s)−len(w)+1) fingerprints. Each fingerprint can be treated as an element in the Sender's set X. At the same time, the label of each element is the set of locations (offsets) of each fingerprint in s. The label may contain multiple locations since a pattern can occur multiple times in the text. This step forms the Sender's set X. The Labelled PSI protocol can be formed for X and Y.

The result from running Labelled PSI gives the matched fingerprints and associated locations in s. Matched patterns and respective locations in s can then be determined. The Receiver can re-roll the window over each of pattern p∈P, and check if all fingerprints from a pattern are matched. If not, then this pattern is not found in text s. If they were all matched, there can be a probabilistic or deterministic implementation. For the probabilistic implementation, the pattern is matched in text with some probability, and the match locations are the locations of the first fingerprint of the pattern. For the deterministic implementation, the locations of these fingerprints are checked to determine if they are consecutively adjacent. If so, then the pattern is found in those locations in the text.

As described herein, one of the optimizations performed for the labelled PSI is to split each of the bins into a smaller bins in order to further reduce the search space per bin; which reduces the degree of the F(x) and G(x) polynomial functions. This optimization is performed at the cost of generating a more F(x) and G(x) polynomial functions and computing a pairs of result ciphertexts instead of a single pair of ciphertexts (one per polynomial) for each query ciphertext. In confidential string-matching, these result ciphertexts are sent back to the Receiver for decryption. In order for the protocol to be practical, the collective size of the results should be smaller than the text corpus s being searched upon. Unfortunately, in order to keep F(x) and G(x) construction practical, the split factor α must scale with the length of s; making the protocol impractical for longer s.

The optimization performed by the system 100 to address the above technical challenge is based on the observation that, advantageously, the a pairs of ciphertext results from the splitting approach contain the same information as the single pair of ciphertext results obtained without splitting. Therefore, it is compressible from an information theory perspective. This stems from the notion that results from both approaches are still solving the same problem. The system 100 takes advantage of the above idea to use an arithmetic circuit that operates on the evaluations of all the F(x) and G(x) (previously known as the result ciphertexts) to reduce the results to a constant number of ciphertext with respect to the length of s.

For match result comparison, in the labelled PSI protocol, for a query x, the function F(x) evaluates to a 0 if the respective coefficient in x matched in the Sender's bin or to an arbitrary value if there is no match. This implies that for each query ciphertext, the set of match ciphertext results can be compressed to a single ciphertext by a multiplication of the results:

U(x)=F₁(x)*F₂(x)* . . . *F_a(x)

where F_i(x) is the evaluation of a F(x) polynomial for the i^th bin in the smaller a set of bins.

For label result comparison, in the labelled PSI protocol, for a query x, the function G(x) evaluates to the label of the respective coefficient in x if it matched in the bin (if the respective coefficient in F(x) is 0); else it evaluates to an arbitrary value. Observe that for a search in a single bin, only one of the items in the respective bin can match as items in a bin are unique. When the bin is split into smaller bins, only one of the smaller bins can match. This implies that for each query coefficient, only one of the respective coefficients from all the F(x) for each bin will be 0. This property can be exploited by the system 100 to reduce the label ciphertext results:

H(x)=G₁(x)*F¹(x)+G₂(x)*F²(x)+ . . . +G_a(x)*F^α(x)

where F^k(x)=Π_i=1,i≠k^a F_i(x) and G_i(x) is the evaluation of a G(x) polynomial for the i^th bin in the smaller a set of bins.

For each query coefficient, the respective coefficient in H(x) will contain an arbitrary value if x did not match in the bin, otherwise it will contain the label for x multiplied by a constant. A separate ciphertext Q(x) is determined, which holds the constant for each query coefficient:

Q(x)=F¹(x)+F²(x)+ . . . +F^α(x)

To extract the label, the multiplicative inverse of decrypted Q(x) is multiplied by the decrypted H(x):

Label=H(x)G⁻¹(x)

FIGS. 7A and 7B illustrate a flowchart diagram of a method for confidential string-matching 200, in accordance with an embodiment. In some cases, the alphabet Σ can consist of 0 and 1 (Σ={0,1}). In other cases, the alphabet can comprise any suitable set of digits (including negative), characters, and/or symbols as part of the alphabet. It is appreciated that at least some of the order and necessity of the steps and limitations of the method 200 can be varied or changed as appropriate, while retaining the technical contributions of the method 200.

At block 202, the first input module 118 receives an input pattern set P⊂Σ*, for example, from the input interface 106, the database 116, or the network interface 114. The input pattern comprises a set of patterns P, where each pattern consists of elements from the alphabet and has minimum length γ.

At blocks 206 to 216, the first pre-processing module 120 performs pre-processing for the Receiver. In some cases, this pre-processing can take place offline, when the first computing device 102 is not necessarily connected to the network 130. At block 206, the first pre-processing module 120 performs windowing by, for all p∈P, determining the hash values for each ω-sized window of p, to form set Y={h₁(w_p)|∀w_p∈p, ∀p∈P}. In most cases, the Receiver and the Sender communicate to agree upon, or use a predetermined value for, a first hash function h₁ which maps an ω-sized window to a value in the range [0, 2^σ) (h₁: {0,1}^w[2^σ]). Thus, for all patterns in pattern set P, the pre-processing module 120 determines the hash values with the first hash function h₁ for each ω-sized window of the pattern, to form the set of hashes Y.

In an example, the co parameter can be selected as less than or equal to the shortest pattern string. In such example, for a specific ω, it can be assumed that there is a window that can hold co characters. The first co characters are selected, then the window is shifted by one or more characters those co characters are formed as part of the window. This is repeated until the end of p is reached.

At block 208, the first pre-processing module 120 performs binning. The parameter m which represents the number of bins, and a second hash function h₂ which maps a value in the range [0, 2^σ) to a value in the range [1,m], are agreed upon to place all items from Y into a table C with m bins. In most cases, the Receiver and the Sender communicate to agree upon, or use a predetermined value for, the parameter m that represents the number of bins, and the second hash function h₂: [2^σ][m], to place all the items from Y into the table C with m bins. In further cases, the second hash function h₂ maps an ω-sized window to the parameter m that represents the number of bins (h₂: {0,1}^w[m]). Such cases can considerably reduce the collision rate and allow for better parameter set selection.

At block 210, the first pre-processing module 120 selects HE parameters. The Receiver and the Sender communicate to agree upon, or use a predetermined value for, parameters (n,q,t) for an indistinguishability under chosen-plaintext attack (IND-CPA) secure HE scheme for the chosen κ and λ. t is selected to be large enough so that each element in Y (represented by σ bits) can be fit into log₂t bits so that log₂t>σ.

At block 212, the first pre-processing module 120 performs batching by interpreting the table C as a 2D array of m rows, where each row represents elements from a single bin. In some cases, for each column of the 2D array, the pre-processing module 120 batches the m elements into plaintext polynomials Y_b. Each plaintext polynomial can be treated as a batch b. While use of the polynomials are not strictly required, they are useful for computational efficiency. In some cases, each ciphertext can encrypt only one element; however, this will be inefficient because the same ciphertext can be used to encrypt a polynomial; the polynomial capable of containing many elements (e.g., tens of thousands of elements).

At block 214, the first pre-processing module 120 determines powers of Y_b. For each Y_b, the pre-processing module 120 determines all the powers of Y_b from 1 to ψ (represented by a set {(Y_b)^k|∀k∈[1,ψ]}), where ψ is the maximum power of Y_b as agreed upon by the Receiver and the Sender.

At block 216, the first pre-processing module 120 performs encryption. For each Y_b, the pre-processing module 120 uses, for example, HE.Encrypt to encrypt each power, (Y_b)^k, to generate ciphertexts: c_b^k=HE. Encrypt((Y_b)^k). At block 218, the network interface 114 communicates the ciphertexts to the Sender.

At block 220, the second input module 168 receives a text corpus s∈Σ*, for example, from the input interface 156, the database 166, or the network interface 164. The text corpus s comprises elements from the alphabet and is of minimum length γ.

At blocks 222 to 230, the second pre-processing module 170 performs pre-processing for the Sender. In some cases, this pre-processing can take place offline, when the second computing device 152 is not necessarily connected to the network 130. At block 222, the second pre-processing module 170 performs windowing. Hash values are computed for each co-sized window of s, to form set of hashes X using agreed upon the first hash function h₁ which maps an ω-sized window to a value in the range [0, 2^σ); i.e., set X={h₁(w_s)|∀w_s∈s} using the agreed upon the first hash function h₁: {0,1}^w[2^σ].

At block 224, the second pre-processing module 170 performs binning. The parameter m and the second hash function h₂ as agreed upon, which maps a value in the range [0, 2^σ) to a value in the range [1,m], h₂: [2^σ][m], to place all items from X into table T with m bins. In further cases, the second hash function h₂ maps an ω-sized window to the parameter m that represents the number of bins (h₂: {0,1}^w[m]). Such cases can considerably reduce the collision rate and allow for better parameter set selection.

At block 226, the second pre-processing module 170 performs splitting. For each bin in T, the second pre-processing module 170 splits it into α sub-bins of size at most ψ, denoted as T[i,1], . . . , T[i,α]; where T[i,j] is the j^th sub-bin in the i^th bin.

At block 228, the second pre-processing module 170 determines coefficients. For each sub-bin T[i,j], the second pre-processing module 170 matches the polynomial by determining a symmetric polynomial F_i,j over F_t such that F_i,j(x)=0 only if x exists in T[i,j]; i.e., x∈T[i,j]. For each sub-bin T[i,j], the second pre-processing module 170 interpolates the polynomial G_i,j over F_t; such that G_i,j(x)=l_x only if x exists in T[i,j]; i.e., x∈T[i,j]. Where l_x is a set of locations {o₁, o₂, . . . o_i} from s where the respective window that hashes to x (w_x) exists.

At block 230, the second pre-processing module 170 performs batching. The polynomials F_i,j are interpreted as a matrix, where i indexes the row and j indexes the column. The second pre-processing module 170 batches the k^th coefficient of each polynomial in a column into one plaintext polynomial F_j,k, which denotes the plaintext polynomial for the k^th coefficient in the j^th sub-bin. The second pre-processing module 170 performs the same batching on the label polynomials G_i,j to form the batched plaintext polynomials G_j,k.

At blocks 232 to 234, the homomorphic evaluation module 172 determines the matching for the Sender homomorphically and compresses the results. At block 232, for the b^th batch received from the Receiver, the homomorphic evaluation module 172 homomorphically evaluates the HE circuit. For each sub-bin, the homomorphic evaluation module 172 homomorphically evaluates the function: f_b,j=F_j,1+Σ_k=1^ψc_b^k−F_j,k+1. The same operation can be repeated for G_b,j and the label ciphertexts g_b,j can be denoted as such.

At block 234, for the b^th batch received from the Receiver, the homomorphic evaluation module 172 performs compression. As described herein, the homomorphic evaluation module 172 compresses f_b,j and g_b,j by evaluating:

u_b=Π_j=1^αf_b,j  (a)

v_b,j=Π_i=1,i≠j^αf_b,i  (b)

q_b=Σ_j=1^αv_b,j  (c)

h_b=Σ_j=1^αg_b,j.  (d)

At block 235, the network interface 164 communicates the match ciphertext u_b, and in some cases, the label ciphertext q_b and the ciphertext h_b to the Receiver. In some cases, u_b and q_b can be determined in a depth-optimal manner by the use of a multiplication tree.

At blocks 236 to 240, the post-processing module 122 determines and outputs a result for the confidential string-matching. In some cases, this post-processing can take place offline, when the first computing device 102 is not necessarily connected to the network 130. At block 236, the post-processing module 122 performs decryption. For the b^th batch, the post-processing module 122 uses, for example, HE.Decrypt on the ciphertext u_b to obtain r_b, which can be interpreted as a vector of n elements in F_t. As such, it forms the corresponding label vector:

l_b=(HE.Decrypt(q_b))⁻¹·HE.Decrypt(h_b)

At block 238, the post-processing module 122 performs query-to-results mapping. For each query vector Y_b, there is a corresponding match vector r_b and label vector l_b. The coefficients from each Y_b vector collectively form the set Y. A table M is formed that is indexed by a coefficient from Y_b and returns a respective coefficient from r_b. Thus, the coefficients from each r_b are used to form a table M such that M[Y_b[i]]=r_b[i] for all batches, where i∈[0,n]. The same approach can be performed to form a table L that is indexed by a coefficient from Y_b and returns a respective coefficient from l_b. Thus, the coefficients from each l_b can be used to form a table L such that L[Y_b[i]]=l_b[i] for all batches.

At block 240, the post-processing module 122 determines the result for the confidential string-matching. For each p∈P, the post-processing module 122 determines hash values for each ω-sized window of p with the first hash function h₁, similar to block 206. Thus, an ordered set H={η=h₁(w₁)|∀w_p∈p} is formed. The pattern p, and the corresponding intersection set O (the locations), are outputted if and only if: (1) the windows from p exist in s (i.e., indexing into M with all hashes in H returns 0); and (2) the windows from p are found located in the same order adjacently in s (i.e., if there exists a subset of locations returned from indexing into L with all hashes in H that are consecutively adjacent with respect to each other). These conditions can be denoted as:

$M\left[{\eta }_i\right]=0\text{∀}{\eta }_i\in H\mathrm{and}\stackrel{N_H}{\bigcap _{i=1}}\mathrm{sub}\left(L\left[{\eta }_i\right],i-1\right)\ne \varnothing$

where sub(S,n)={s−n|∀s∈S}.

Thus, for each query polynomial, the post-processing module 122 determines a match polynomial and, in some cases, a label polynomial. Each of the coefficients generally represent an independent query (search for respective fingerprints). If the respective fingerprint matched, then the corresponding coefficient in the match polynomial will be zero; and the corresponding coefficient in the label polynomial will contain a list of locations where such fingerprint was located. In order to link the window-fingerprint search to the overall pattern search, tables M and L are used, which contain the match results and label results, respectively, for all the query polynomials. Then for each pattern, the post-processing module 122 determines the pattern's respective window-fingerprints and checks the results for the window-fingerprints in M and L. If the window-fingerprints are all 0 in M (representing that they all were found), then the post-processing module 122 checks if these windows exist adjacent to each other in s, in the same order they are in the respective pattern, using locations from L for each window in the pattern. Hence, the post-processing module 122 checks if there is any location where the location of the first window equals the location of the second window minus one, which equals the location of the third window minus two, and so on, for all the windows of the respective pattern.

In a further embodiment, the system 100 can be used for a type of confidential string-matching referred to as confidential deep packet inspection (DPI). Generally, DPI is a type of data processing that inspects, in detail, the data being sent over a computer network. In many approaches, DPI consists of searching for signatures in a network data stream and executing rules based on the results. It is beneficial to be able to perform DPI on some network traffic in an insecure environment where the signatures being searched for and the results must remain confidential. Such a system should ideally be able to perform DPI in a practical manner at high network speed. These requirements present a substantial technical challenge and can be referred to as the confidential DPI problem.

Advantageously, the system 100 can be used to substantially address the security and efficiency constraints of the confidential DPI problem; for example, by transforming the string-matching problem described herein into the confidential DPI problem. In the confidential DPI problem, the first computing device 102 (in this embodiment, referred to as an Agency) holds a set of signatures to be kept private, and the second computing device 152 (in this embodiment, referred to as a Host) receives a continuous stream of packets. This stream is segmented according to a fixed time-interval, and DPI is performed on each segment. DPI on each segment is mapped to the confidential string-matching protocol, described herein, to leverage its performance and efficiency (where the Agency is the Receiver and the Host is the Sender). Here the Agency is in a “Secure Environment” and the Host may be in an “Insecure Environment”. In practice, this embodiment has been determined to be able to process many signatures (e.g., tens of thousands) against a high-speed network (e.g., Gbps). This embodiment can have many applications, for example:

• • Critical Infrastructure Monitoring: Governments can monitor the networks of their critical infrastructures (e.g., mobility, power, nuclear, communication, water supply, and the like) from attacks coming from suspicious or threatening sources (e.g., hackers, dark web, nation state attacks, and the like). Signatures identifying these sources can be encrypted and placed in network nodes inside the critical infrastructure networks for continuous monitoring and intrusion detection.
  • IP Protection: Confidential DPI can be integrated into network boundary firewalls to be aware of Intellectual Property leaving the protected network (data exfiltration).
  • Anti-terrorism and Dark Web Surveillance: An agency can perform inspection on a network without revealing sensitive signatures or results to the public.
  • Money Laundering Monitoring: An agency can perform monitoring on networks suspected of money laundering traffic.
  • Insider Trading Monitoring: An agency can perform monitoring on private networks suspected of insider trading traffic without exposing the data on the private networks (e.g., a government can mandate DPI on internal networks of technology companies).

The confidential string-matching problem, described above, and the confidential DPI problem share many similarities. They search for a set of patterns in a collection of strings. Described herein is how the confidential string-matching problem can be mapped to the confidential DPI problem. FIG. 8 illustrates a diagram of an example high-level overview of an approach to addressing the Confidential DPI problem by the system 100.

The Agency's signature set Z⊂Σ* can be directly treated as the pattern set P in confidential string-matching, described herein. Observe that instead of a text string s, the DPI problem has a continuous stream of packets, where each packet contains a packet body and its packet 5-tuple. This stream can be segmented using a time interval. The problem can be resolved for each segment one-by-one. As the stream of packets is continuous, the collection of the segments can be pipelined with the determination on collected segments. For each segment, instead of one string to search in the confidential string-matching protocol, in confidential DPI there will be multiple strings to search. Therefore, all packet bodies are fingerprinted across all packets in the segment.

In the confidential DPI problem, the Agency needs to know which signature matched which packet (5-tuple) and where it matched within the packet (packet offset). This implies that for each query, the Host needs to return the set of packet 5-tuple and offsets where the query matched. The packet offsets can be encoded into the label ciphertext in the same way as the location in the confidential string-matching problem. However, in an example, in Internet Protocol version 4 (IPv4), a single packet 5-tuple can be 13 bytes; which in the HE approach used in the confidential string-matching problem requires multiple (e.g. four) ciphertext coefficients to store, leading to additional multiple label ciphertexts per match for each query ciphertext. Also, since the system 100 does not know how many matches will happen, it can be assumed that there are the maximum matches, and that this many label ciphertexts can be sent. In some cases, to avoid sending a large quantity of label ciphertext, the Host sends back 5-tuples for all packets unencrypted. The label ciphertexts can, in addition to the packet offsets, contain a set of IDs to identify each matched packet. Once the Agency decrypts the label ciphertext, it can use the ID to obtain the 5-tuple from the packet 5-tuple list. Therefore, during interpolation, each location in the label for a fingerprint is of the form ω={ID_packet∥offset}. Note that concatenation of the ID and offset in this order allows the adjacency checking in the string-matching system to function unmodified; i.e., the system 100 can check if two windows are adjacent to each in the same packet. After the Host performs string-matching, the Agency obtains a list of matched signatures and associated packet IDs and offsets. The packet IDs can be converted to packet 5-tuples by looking them up in the list of 5-tuples.

The same optimizations applied in the confidential string-matching, described above, can be applied in the confidential DPI; for example, data compression. In some cases, to satisfy the real time processing requirement for DPI, pipelining can be used to break the computation process into multiple stages, where each stage has a full time window for its own processing. This allows the overall process to be fitted within the specified time window.

FIGS. 9A and 9B illustrate a flowchart diagram of a method for confidential deep packet inspection using confidential string-matching 900, in accordance with an embodiment. In some cases, the alphabet Σ can consist of 0 and 1 (Σ={0,1}). In other cases, the alphabet can comprise any suitable set of digits (including negative), characters, and/or symbols as part of the alphabet.

At block 902, the first input module 118 receives an input signature set Z⊂Σ*, for example, from the input interface 106, the database 116, or the network interface 114. The input pattern comprises a set of signatures Z, where each signature consists of elements from the alphabet Σ and has minimum length γ. Signatures can also be referred to as patterns, as in the method 200.

Similar to in method 200, except replacing P with Z, the first pre-processing module 120 performs windowing at block 904, the first pre-processing module 120 performs binning at block 906, the first pre-processing module 120 selects HE parameters at block 908, the first pre-processing module 120 performs batching at block 910, the first pre-processing module 120 determines powers at block 912, and the first pre-processing module 120 performs encryption at block 914 to determine encrypted ciphertexts. In some cases, this pre-processing can take place offline, when the first computing device 102 is not necessarily connected to the network 130. At block 915, the network interface 114 communicates the ciphertexts to the second computing device 152.

At block 916, the packet module 174 buffers the network packets within a specified time interval to form a list of packets D={(a,b)|a⊂Σ*, b⊂Σ*}; where each element in D is an ordered pair of packet body (a) and packet 5-tuple (b). The packet body and the 5-tuple consists of elements from the alphabet Σ and have a minimum length γ. After a complete time interval passes, the packet module 174 forwards the collected packets to be processed and starts buffering the packets in the next time interval.

At block 918, the packet module 174 performs windowing. For all packet bodies in the packets D, it uses Rabin-Karp to hash each ω-sized window of the packet bodies, to form set of hashes X using agreed upon first hash function h₁ which maps an ω-sized window to a value in the range [0, 2^σ). In this way, for all a∈D_a, it uses Rabin-Karp to hash each ω-sized window of a, to form set X={h_i(w_d)|∀w_a∈a, ∀a∈D_a} using agreed upon first hash function h₁:{0,1}^w[2^σ].

At block 920, the packet module 174 communicates all packet 5-tuples for the packets D_b to the Agency.

Similar to in method 200, the second pre-processing module 170 performs binning at block 924 and the second pre-processing module 170 performs splitting at block 926. In some cases, this pre-processing can take place offline, when the second computing device 152 is not necessarily connected to the network 130.

Similar to in method 200, the second pre-processing module 170 determines coefficients at block 928, except in this embodiment, the second pre-processing module 170 labels polynomials differently. The second pre-processing module 170 interpolates the polynomial G_i,j over F_t such that G_i,j(x)=l_x only if x exists in T[i,j] (G_i,j(x)=l_x∀x∈G[i,j]). Where l_x is a set of unique locations for windows that hash to x (w_x) across all packet bodies, D_a. The unique locations are represented by the concatenation ω={ID_packet∥offset}. Where ID_packet is the unique identifier of a packet in D, and the offset is the position, of the window that hashes to x, within the packet body. In other words, ID_packet is the unique identifier of a packet in D, and offset is the position of the w_x within the packet body.

Similar to in method 200, the second pre-processing module 170 performs batching at block 930. Also similar to method 200, the homomorphic evaluation module 172 determines intersection by, for the b^th batch received, homomorphically evaluating the HE circuit at block 932. For each sub-bin, the homomorphic evaluation module 172 homomorphically evaluates the function: f_b,j=F_j,1+Σ_k=1^ψc_b^k·F_j,k+1. And performs the same operation for G_b,j and the label ciphertexts g_b,j are denoted as such. At block 936, for the b^th batch received, the homomorphic evaluation module 172 performs compression. Similar to in method 200, at block 937, the network interface 164 communicates the match ciphertext u_b, the label ciphertext q_b, and the ciphertext h_b to the Agency. In some cases, u_b and q_b can be determined in a depth-optimal manner by the use of a multiplication tree.

Similar to in method 200, except replacing P with Z, the post-processing module 122 performs decryption at block 938 and the post-processing module 122 performs query-to-results mapping at block 940.

At block 942, the post-processing module 122 determines the result for the confidential DPI. For each z E Z, block 940 will return a ω={ID_packet∥offset} for every location where z matched in the packet bodies of D. For each ω, the post-processing module 122 retrieves the 5-tuple from the list of 5-tuples received earlier using the ID_packet. In some cases, the above post-processing can take place offline, when the first computing device 102 is not necessarily connected to the network 130.

At block 944, the post-processing module 122 outputs the 5-tuple and the offset. The output includes D_a∩Z and the 5-tuple D_b corresponding to it, as well as the offsets of these signatures within the packets. In this way, the Agency outputs the signatures that are found in any of the packet bodies, the corresponding 5-tuple for those packets, as well as the offsets within the packets where the signatures matched.

In further cases, the system 100 can be used in, for example, network Intrusion Detection System (IDS), Intrusion Prevention System (IPS), and Network Security Monitoring (NSM) engine (e.g. Snort and Suricata).

In further embodiments, the system 100 can be used for regular expression (RegEx) patterns by windowing the pattern around the RegEx characters, finding those patterns and their respective offsets, and performing some post-decryption logic using the offsets to find if the RegEx matched. For a regular expression to be supported, each contiguous set of non-RegEx characters in a pattern is expected to be equal to or larger than the window size.

In an example for the “.” wildcard:

• • RegEx Pattern: “abc.efg”
  • Text Body: “xyzabcyefgzxyz”
  • Window Size ω: 3
    A search can be performed for “abc” and “efg” resulting in offset 3 and 7 respectively. If “abc” has offset x and “efg” has offset x+4, then the RegEx pattern is satisfied.

In another example for the “.*” wildcard quantification:

• • RegEx Pattern: “abc.*efg”
  • Text Body: “xyzabcyxzyb53dfxfefgzxyz”
  • Window Size ω: 3
    A search can be performed for “abc” and “efg” resulting in offset 3 and 17 respectively. If “abc” has offset x and “efg” has offset x+k, where k>=ω, then the RegEx pattern is satisfied.

In another example for the “?” quantifier:

• • RegEx Pattern: “abcy?efg”
  • Text Body: “xyzabcyefgzxyz”
  • Window Size ω: 3

A search can be performed for “abc”, “bcy”, “efg” resulting in offset 3, 4 and 7 respectively. If “abc” has offset x and only if the following conditions is true: (1) “efg” has offset x+4 and “bcy” has an offset x+1, and (2) “efg” has offset x+3, then the RegEx pattern is satisfied.

In another example for the “*” quantifier:

• • RegEx Pattern: “abcy*efg”
  • Text Body: “xyzabcyyyyyyyyyyefgzxyz”
  • Window Size ω: 3

A search can be performed for “abc” and “efg” resulting in offset 3 and 17 respectively. A search should also be performed for the possible windows that would result in the RegEx pattern match, namely: “bcy”, “cyy”, “yyy”, “yye”, “yef”. Using the offsets for these windows from the search, it can be extrapolated, the information about the contents in the text body between any “abc” and “efg” windows; and hence, determine if the RegEx pattern is satisfied. The “+” as well as the “{n}”, “{min,}” and “{min, max}” quantifiers can be handled in a similar way.

In another example for the “I” construct for “or” with grouping “Q”:

• • RegEx Pattern: “abc(mno|pqrs)efg”
  • Text Body: “xyzabcmnoefgzxyz”
  • Window Size ω: 3

A search can be performed for {“abc”, “efg”} as well as the windows in grouping {“mno”, “pqr”, “qrs”} resulting in offset 3, 9, 4, ∅, ∅ respectively. If “abc” has offset x and one of the following is true: (1) “mno” has offset x+1 and “efg” has offset x+k, where k==len(“mno”), and (2) “pqr” has offset x+1, “qrs” has offset x+2 and “efg” has offset x+j, where j==len(“pqrs”), then the RegEx pattern is satisfied.

In another embodiment, the system 100 can remove the windowing on the patterns. Instead of searching for fingerprints of the windows from a pattern, the system 100 can search for the fingerprint of the whole pattern. In this case, the system 100 can generate a single fingerprint for each pattern. The system 100 can also generate a fingerprint for each substring in the text corpus that has the potential to match with any of the patterns. This implies that a range is to be defined for the possible pattern length [γ,υ], and translates to creating fingerprints for all substrings in the text corpus that have length in the range [γ,υ].

This above approach allows for exact matches (assuming negligible collision probability) of the patterns without the need for offsets; as now there is no need to verify adjacency of the windows from a pattern as in the prior. It also allows for smaller label size because the system 100 is now matching fingerprints of a pattern. The label will indicate the locations of the full pattern as opposed to locations of a window from the pattern. This translates to potentially less locations as window sizes are smaller than the patterns; and therefore, have higher probability for matching. This also means generally no false negatives due to the scenario where all matched locations for the pattern fit in the label (of pre-determined size). Whereas in the windowing approach, the matched locations for a window of the pattern may not fit in the label. However, this can come at the cost of N_x, the size of the set of fingerprints from text corpus, increasing with the pattern length range; i.e., N_x ∝(υ−γ). Therefore, this approach is particularly apt when the pattern length range is small and the bandwidth for the label ciphertext is limited. This approach is also particularly applicable to confidential DPI, as described herein, with a fingerprint generated and searched, for each signature.

In another embodiment of string-matching, the patterns can be windowed in a different way. Instead of fingerprinting each possible ω-sized substring in a pattern, the system 100 only fingerprints the non-overlapping ω-sized substrings in the pattern; i.e., the pattern is segmented into ω-sized substrings. For patterns where the pattern size is not divisible by ω, the last window of the pattern can overlap. This is advantageous for scenarios where window size is much smaller than average pattern length as it substantially reduces the Receiver's fingerprint set Y size; i.e., it reduces the number of windows to search for. In this embodiment, if the offsets are not communicated (i.e., the probabilistic result interpretation is used), then the probability of getting a false positive (the windows of a pattern exists in the text corpus, but the pattern itself does not) is substantially higher. The aspects of this embodiment can also be compatible with a variation of the confidential DPI embodiment.

As an example, the pattern “abcdef”, using the embodiment of method 200, the windows will be:

• • “abc”
  • “bcd”
  • “cde”
  • “def”
    If the pattern exists in the text corpus, offsets for the window will be x, x+1, x+2 and x+3 respectively.

In the above embodiment, the system 100 only needs to search for non-overlapping windows:

• • “abc”
  • “def”
    If the pattern exists in the text corpus, offsets for the windows will be x and x+3 respectively; which fully defines that “abcdef” existed in a text corpus at location x.

In another embodiment, an R-extension can be used for the string-matching. For the probabilistic implementation described herein, the probability of getting false positives (cases where a match is falsely reported for a pattern that does not exist in the text corpus) can be substantially reduced by incorporating window adjacency information into the query. Doing so will result in the Sender effectively searching for the full pattern instead of substrings from the pattern.

False positives generally come from the loss of adjacency information between windows when the patterns (and text corpus) are windowed and fingerprinted into a set X (& Y). In this embodiment, the window adjacency information can be incorporated by foregoing the formation of set X and Y and combining the windowing and binning as follows:

• • As the window is rolled across the patterns, for each window (w_p), the window's fingerprint (h₁(w_p)) is inserted into the first table at location h₂(h₁(w_p)) and the successive window's fingerprint (in the pattern) is inserted in each subsequent table at the same location h₂(h₁(w_p)). Therefore, the first table is identical to the table described with respect to the other embodiments, while the other tables contain information about (R−1) windows that are adjacent to windows with fingerprints in the first table.
  • Instead of a single ciphertext representing a query batch, a query batch is represented with R ciphertexts as there are R tables instead of a single table.
  • The same windowing and binning can be applied on the text corpus to create R tables with fingerprints of windows from the text corpus. In this variation, the system 100 can then attempt to match both the windows and their adjacent windows by mapping the R query ciphertexts to the R respective tables. The results from the search on each table are added together such that a match only occurs if both the window and its adjacent windows are found in the text corpus. Thus, for a pattern containing exactly R windows, the probability of false positives is significantly reduced.
  • For the case where a window does not have R−1 adjacent windows in the pattern, zeroes can be used to denote “no window” in the subsequent tables. A zero can be added to every bin on the text corpus side (before alpha splitting) so that a search for zero in any of the bins will result in a match; and therefore, will not factor into the addition of the results. Thus, translating to a significant reduction in the probability of false positives for patterns with less than R windows.
  • For the case where a pattern has more than R windows, the results from each set of R (or less) windows from the pattern can be combined together. The overall probability of false positives will be lower compared to the method 200 but higher compared to the case where the pattern has R or less windows.
  • If the location of the patterns in the text corpus is required, the embodiment of method 200 using labels remains unchanged; i.e., the system 100 only needs to evaluate the label results for the first ciphertext in the R ciphertexts batch. The G (x) interpolation polynomial is not constructed for any bins in the tables besides the first table on the text corpus side.

FIG. 10 illustrates an example of the R-extension embodiment. FIG. 10 shows an example of confidential string-matching with R-extension using h₁ for fingerprinting windows and h₂ for binning the fingerprints to form query ciphertexts and tables to search. The pattern “abcde” is windowed across the R query ciphertexts so that the results for the search of “abc”, “bcd” and “cde” can be added together; only returning a match if they were adjacent in the text corpus. The figure also highlights that the pattern “deab” is not found even though the windows “dea” and “eab” exist separately in the text corpus. Note that the “*” character in the figure denotes an arbitrary non-zero value, which depending on the context, can indicate “no match” or “don't care”.

As shown in FIG. 10, the pattern “abcd” occupies three rows across the query ciphertext to find if “abcde”, “bcde” and “cde” exist in the corpus respectively. From a matching perspective, this is redundant and the rows to search for “bcde” and “cde” can be removed as the search for the pattern is complete with just the first row. However, when the location of the pattern is also needed (achieved via labels), the other rows will be required as the first row will return all locations where “abc” exists instead of only the locations where “abcde” exist. Combining the locations from “bcde” and “cde” will give the locations where only the pattern exists.

In some cases, the text body (or network traffic in case of DPI) can be treated confidential as well. This can be very useful in the case that the text body and the queries are encrypted using the same key pair (public and secret keys). For example, a bank encrypts its sensitive databases and uploads the encrypted contents to the cloud to allow the bank agents to perform encrypted queries on such databases. Additionally, this allows for cloud migration strategies.

In some cases, the system 100 only requires knowing if a pattern exists in a corpus, not its location or any other metadata associated with its existence. This will eliminate the “Labelled” function altogether and use only the function F(x).

In some cases, the system 100 can support binary operations (e.g., AND, OR, and the like) between multiple signatures by performing these steps as a post-processing step.

In some cases, for confidential DPI, the system 100 can enable signature flags based on signature offset within the packet by taking advantage of the packet offset information retrieved by the label.

In some cases, other techniques (e.g., Chinese Remainder Theorem (CRT) and Single Instruction Multiple Data (SIMD)) can be used to break down the computation on wide integers (for example hundreds or even thousands of bits) into multiple computations over narrower numbers that can fit into computer registers and the processor arithmetic logic unit (for example 64 or 128 bits).

In some cases, the system 100 can make use of a fixed-sized time interval buffering to enable processing on a continuous stream. The time interval can be variable-sized and adjusted dynamically based on the statistical properties of the input traffic. Additionally, if the input is non-continuous, time interval buffering may not be required.

In some cases, approaches described herein can be less efficient in terms of communication cost, as the number of result ciphertexts increase by a factor of α. Instead, the system 100 can remove the factor α and have larger bins. However, this will increase the interpolation time. This approach can be useful for the case that the text corpus being searched is static. In this case, interpolation can be performed offline.

In some cases, parallelization and/or pipelining can be not performed. In these cases, the approach may take longer to run, but can allow for running on more platforms (e.g., internet-of-things, embedded, and mobile processors).

In some cases, the labels can be communicated using:

• • Private Information Retrieval (PIR): In the case that the desired “labels” are large, for instance files, an alternative approach can be taken. Only the existence of a match is determined along with a unique ID corresponding to the desired “label” using the PSI routine described herein. Note that the unique IDs used should be much smaller than the desired “label”. Then, the unique IDs are utilized as keys to retrieve the “label” using an HE private information retrieval (PIR) routine. This approach will be generally more efficient in regards to computation and communication when the “labels” are all similarly large in size.
  • Oblivious Pseudo Random Function (OPRF): An OPRF can be applied to both the signature windows and the text corpus windows. The OPRF processing can be performed using exponentiation or an oblivious transfer to hash the items on both sides. Additionally, the labels can be encrypted by applying the OPRF hash function. All the encrypted labels can then be communicated back to the Agency as they will be pseudo-random in the Agency's view. Sending all the labels will generally increase the communication cost; however, it is also possible to combine the OPRF technique with the PSI routine.

In both method 200 and method 900, there is polynomial creation (expansion) and interpolation. There are several approaches that can achieve this, each with different space and time complexities. It is appreciated that at least some of the order and necessity of the steps and limitations of the method 900 can be varied or changed as appropriate, while retaining the technical contributions of the method 900.

In one such approach for polynomial creation, to create the polynomial using convolution, several smaller convolution operations are performed using all the roots of the polynomial. Each root is placed in an array of size 2 where index 0 is a 1 and index 1 is the negative of the root in the finite field space. Then all the arrays are convolved in sequence.

ArrayofCoefficients=[1,−r₁]*[1,−r₂]*[1,−r₃]* . . . *[1,−r_n]

The time complexity of the convolution is O(n²).

In another such approach for polynomial creation, Vieta's algorithm uses the sum of roots and sum of products equations and generalizes them to find all the coefficients of the polynomial. The time complexity of Vieta's algorithm is also O(n²) but its constant factor is smaller, leading it to be faster in practice.

In another such approach for polynomial creation, number theoretic transform (NTT) works by transforming the roots of the polynomial into a different space so that the convolution operation is cheaper, making it equivalent to multiplication. This results in massive speedups in time with an overhead cost of converting to and from this space. The time complexity of the NTT algorithm is O(n·log (n)).

In another such approach for polynomial creation, Kronecker substitution (KS) can be used to treat each array in the convolution as its own second degree polynomial. KS can be used to multiply the polynomials extremely quickly for all pairs of polynomials to generate n/2 third-degree polynomials. This process is repeated with all the third-degree polynomials to generate fifth degree polynomials, and so on, until the coefficients of the final n^th degree polynomial are generated. The time complexity of the best performing KS algorithm is O(n·log (n)) but it has been shown to be faster than NTT up to polynomials of certain degree (e.g., polynomial degree 2¹³).

For polynomial interpolation, a fast implementation is beneficial to realize a practical, high performance implementation. The goal of polynomial interpolation is to fit an n^th order polynomial to n data points. It can be shown that such a polynomial exists and that there is only one polynomial that exactly fits those points. As described below, provided are examples of different approaches for interpolating a polynomial given a set of x and y points.

In one such approach for polynomial interpolation, a Vandermonde matrix can be used. This approach consists of constructing a linear system of equations and solving them. The system of equations can include:

Xa=Y

where X and Y are known. The vector a can be determined by inverting X:

a=X⁻¹Y

This gives us the coefficients for the interpolating polynomial however the time complexity of matrix inversion is O(n³).

In another such approach for polynomial interpolation, a Lagrange approach can be used. The Lagrange approach of interpolation involves creating basis polynomials L_i(x). Each basis polynomial is constructed so that it goes through 0 for all of the x values except the k^th, for which it has the value 1. Each basis polynomial is then scaled by y, so the basis polynomials go through y for the point in x they correspond to, and zero for all the other points. These polynomials are then added to get the final result. The runtime of this algorithm is O(n²).

In another such approach for polynomial interpolation, a Newton approach can be used. The Newton approach of polynomial interpolation relies on computing “divided differences”. It is similar to the Lagrange approach in that it is a linear combination of basis functions, but in this case the basis functions are Newton polynomials. The runtime of this algorithm is also O(n²).

In another such approach for polynomial interpolation, fast polynomial interpolation over finite field can be used. Fast implementations of the polynomial interpolation algorithm take advantage of the mod space in finite fields. An example of the fast Lagrange approach can be used, which relies on fast polynomial multiplication and division. In addition, it also uses the divide and conquer technique. The complexity of this algorithm is O(n(log n)²). A sub-product tree of weights can be constructed efficiently in order to compute the Lagrange basis functions. The implementation is further optimized by storing and reusing the computed Lagrange basis functions, whenever the set of x values remains constant and only they values are updated.

In some cases, permutation based hashing can be used for method 200 and method 900. The length of the items that can be encoded into a coefficient of an HE ciphertext is generally limited by the value of the plaintext modulus t. To mitigate this limitation, permutation based hashing can be applied on the items stored in the hash table. This technique compresses the bit string of an item by encoding a portion of its bits into the bin location. The hashing can be applied to signature windows, and packet windows as part of the binning; and consequently, only the compressed windows are encrypted in the ciphertext. The size of the bit length reduction achieved with this hashing technique depends on the number of bins used in the hash table.

While the above embodiments generally describe communication between two computing devices, such as the first computing device 102 and the second computing device 152, it is understood that the functions of the system 100 can take place on a single computing device. For example, for performing confidential string matching of method 200 on an encrypted database on the computing device.

In further embodiments, if there is a very large database of patterns P to be searched, the first preprocessing module 120 and/or the second preprocessing module 170 can split up the database into multiple “subsets”. Each subset can be treated as a completely independent database to be searched. This advantageously allows for parallelization of the different subsets on different threads or different systems. Additionally, in an example where there are k subsets, and there is an item that cannot be added to any of the subsets due to collisions as a result of hashing (using the first hash function h₁), the first preprocessing module 120 and/or the second preprocessing module 170 can insert this item in subset k+1 (i.e., create a new subset) which can be used to hold colliding items.

Advantageously, embodiments of the present disclosure leverage homomorphic encryption (HE) to satisfy security constraints of the confidential string-matching problem. The compression technique can be applied in association with hardware acceleration techniques to transform the difficult technical problem into a practical solution.

Although the invention has been described with reference to certain specific embodiments, various modifications thereof will be apparent to those skilled in the art without departing from the spirit and scope of the invention as outlined in the claims appended hereto. The entire disclosures of all references recited above are incorporated herein by reference.

Claims

1. A method for confidential string-matching of input patterns to a corpus, each input pattern comprises elements from an alphabet, the corpus comprises elements from the alphabet, the method executed on at least one processing unit of a second computing device, the method comprising:

receiving encrypted ciphertexts from a first computing device;

windowing the corpus and applying a hash function on the corpus windows to form a corpus set of hashes;

performing binning on the corpus set of hashes;

performing batching on the binned corpus set of hashes;

determining match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes; and

communicating the match ciphertexts to the first computing device to determine the confidential string matching.

2. The method of claim 1, wherein the first computing device determines the confidential string matching by:

decrypting the match ciphertexts, and

determining, from the output of the decryption, if the hash value for each pattern window matches the hash value for any corpus windows and if the matched windows are adjacent in the corpus.

3. The method of claim 1, wherein the first computing device having determined the encrypted ciphertexts by:

determining a hashed set from windowing the input patterns and determining hash values for each pattern window;

performing binning by placing items in the hashed set into a table;

batching the elements of the table into polynomials; and

performing encryption on the polynomials to generate the encrypted ciphertexts;

4. The method of claim 1, wherein performing the binning comprises mapping all elements in the corpus set of hashes to one of a plurality of bins, and splitting the elements in each of the bins into one of a plurality of sub-bins.

5. The method of claim 4, wherein performing the batching comprises determining coefficients of a polynomial for each sub-bin, and batching the coefficients of each polynomial into a plaintext polynomial.

6. The method of claim 5, wherein evaluating the homomorphic encryption circuit comprises homomorphically evaluating a function comprising a combination of the ciphertexts and the plaintext polynomials associated with each of the sub-bins.

7. The method of claim 6, further comprising performing compression on the evaluated function.

8. The method claim 6, wherein the homomorphic encryption comprises an indistinguishability under chosen-plaintext attack (IND-CPA) secure homomorphic encryption scheme, and wherein the parameters for the IND-CPA secure homomorphic encryption scheme are communicated between the second computing device and the first computing device.

9. The method of claim 1, wherein evaluating the homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes further determines a label ciphertext that encodes a set of these locations of the match ciphertexts.

10. A system for confidential string-matching of input patterns to a corpus, each input pattern comprises elements from an alphabet, the corpus comprises elements from the alphabet, the system comprises at least one processing unit and a data storage media in communication with the at least one processing unit, the at least one processing unit and the data storage media part of a second computing device, the at least one processing unit configured to execute:

a second input module to receive encrypted ciphertexts from a first computing device;

a second pre-processing module to window the corpus and applying a hash function on the corpus windows to form a corpus set of hashes, to perform binning on the corpus set of hashes, and to perform batching on the binned corpus set of hashes;

a homomorphic evaluation module to determine match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes; and

a network interface to communicate the match ciphertexts to the first computing device to determine the confidential string matching.

11. The system of claim 10, wherein the first computing device comprises at least one processing unit, and wherein determining the confidential string matching is performed by executing on the at least one processing unit:

decrypting the match ciphertexts, and

determining, from the output of the decryption, if the hash value for each pattern window matches the hash value for any corpus windows and if the matched windows are adjacent in the corpus.

12. The system of claim 11, wherein the first computing device having determined the encrypted ciphertexts by executing on the at least one processing unit:

determining a hashed set from windowing the input patterns and determining hash values for each pattern window;

performing binning by placing items in the hashed set into a table;

batching the elements of the table into polynomials; and

performing encryption on the polynomials to generate the encrypted ciphertexts;

13. The system of claim 10, wherein performing the binning comprises mapping all elements in the corpus set of hashes to one of a plurality of bins, and splitting the elements in each of the bins into one of a plurality of sub-bins.

14. The system of claim 13, wherein performing the batching comprises determining coefficients of a polynomial for each sub-bin, and batching the coefficients of each polynomial into a plaintext polynomial.

15. The system of claim 14, wherein evaluating the homomorphic encryption circuit comprises homomorphically evaluating a function comprising a combination of the ciphertexts and the plaintext polynomials associated with each of the sub-bins.

16. The system of claim 15, wherein the homomorphic evaluation module further performs compression on the evaluated function.

17. The system of claim 15, wherein the homomorphic encryption comprises an indistinguishability under chosen-plaintext attack (IND-CPA) secure homomorphic encryption scheme, and wherein the parameters for the IND-CPA secure homomorphic encryption scheme are communicated between the second computing device and the first computing device.

18. The system of claim 10, wherein evaluating the homomorphic encryption circuit between the encrypted ciphertexts and the batched corpus set of hashes further determines a label ciphertext that encodes a set of these locations of the match ciphertexts.

19. A method for confidential deep-packet inspection of input signatures to traffic over a network, the input signatures comprise elements from an alphabet, the method executed on at least one processing unit of a second computing device, the method comprising:

second input module receiving encrypted ciphertexts from a first computing device, the first computing device having determined the encrypted ciphertexts by: determining a hashed set from windowing the input signatures and determining hash values for each signature window, performing binning by placing items in the hashed set into a table, batching the elements of the table into polynomials, and performing encryption on the polynomials to generate the encrypted ciphertexts;

buffering the network traffic in a specified time interval to form a list of packets, each packet comprising a packet body and a packet 5-tuple, the packet body and the packet 5-tuple comprise elements from the alphabet;

windowing each packet body and applying a hash function on the packet windows to form a packet set of hashes;

performing binning and splitting on the packet set of hashes;

performing batching on the binned and split packet set of hashes;

determining match ciphertexts by evaluating a homomorphic encryption circuit between the encrypted ciphertexts and the batched packet set of hashes; and

communicating the match ciphertexts and the packet 5-tuples to the first computing device, the confidential deep-packet inspection determinable by the first computing device by: decrypting the match ciphertexts, determining, from the output of the decryption, if the hash value for each signature window matches the hash value for any packet windows, determining if the matched windows are adjacent in a packet, and determining the associated 5-tuple from the received packet 5-tuples.

20. The method of claim 19, wherein performing the binning comprises mapping all elements in the corpus set of hashes to one of a plurality of bins, and splitting the elements in each of the bins into one of a plurality of sub-bins, wherein performing the batching comprises determining coefficients of a polynomial for each sub-bin, and batching the coefficients of each polynomial into a plaintext polynomial, and wherein evaluating the homomorphic encryption circuit comprises homomorphically evaluating a function comprising a combination of the ciphertexts and the plaintext polynomials associated with each of the sub-bins.