Abstract: Privacy-preserving homomorphic inferencing utilizes batch processing on encrypted data records. Each data record has a private data portion of interest against which the inferencing is carried out. Batch processing is enabled with respect to a set of encrypted data records by techniques that ensure that each encrypted data record has its associated private data portion in a unique location relative to the other data records. The set of encrypted data records are then summed to generate a single encrypted data record against which the inferencing is done. In a first embodiment, the private data portions of interest are selectively and uniquely positioned at runtime (when the inferencing is being applied). In a second embodiment, the private data portions of interest are initially positioned with the data-at-rest, preferably in an off-line process; thereafter, at runtime individual encrypted data records are processed as necessary to adjust the private data portions to unique positions prior to batching.

Abstract: A computer-implemented method, system and computer program product for utilizing a variational autoencoder for neighborhood sampling. A variational autoencoder is trained to generate in-distribution neighborhood samples. Upon training the variational autoencoder to generate in-distribution neighborhood samples, in-distribution neighborhood samples of an instance of a dataset in latent space that satisfy a distortion constraint are generated using the trained variational autoencoder. A set of interpretable examples for the in-distribution neighborhood samples are then generated using a k-nearest neighbors algorithm. Such interpretable examples are then used to explain the black box model's predictions. As a result, the accuracy of the decision making ability of post-hoc local explanation methods is improved.

Abstract: An example operation may include one or more of generating, by a training participant client, a plurality of transaction proposals, each of the plurality of transaction proposals corresponding to a training iteration for machine learning model training related to stochastic gradient descent, the machine learning model training comprising a plurality of training iterations, the transaction proposals comprising a gradient calculation performed by the training participant client, transferring the plurality of transaction proposals to one or more endorser nodes or peers each comprising a verify gradient smart contract, executing, by each of the endorser nodes or peers, the verify gradient smart contract; and providing endorsements corresponding to the plurality of transaction proposals to the training participation client.

Abstract: A computer-implemented method, system and computer program product for generating in-distribution samples of data for a neighborhood distribution to be used by post-hoc local explanation methods. An autoencoder is trained to generate in-distribution samples of input data for the neighborhood distribution to be used by a post-hoc local explanation method. Such training includes mapping the input data (e.g., time series data) into a latent dimension (or latent space) forming a first and a second latent code. A mixed code is then obtained by convexly combining the first and second latent codes with a random coefficient. The mixed code is then decoded with the input data masked with interpretable features to obtain conditional mixed reconstructions. Adversarial training is then performed against a discriminator in order to promote in-distribution samples by computing the reconstruction losses of the conditional mixed reconstructions as well as the discriminator losses and then minimizing such losses.

Abstract: An example operation may include one or more of generating a hashed summary including hashes of one or more of a validation data set and hashes of data points chosen in previous iterations from producer nodes, and exposing the hashed summary to a plurality of producer nodes, receiving, iteratively, a plurality of requests from the plurality of producer nodes, respectively, where each request identifies a marginal value provided by a hash of a data sample available to a producer node, selecting a request received from a producer node based on a marginal value associated with the request, retrieving hashed data of the producer node associated with the selected request, and aggregating the hashed data of the producer node with the summary of hashes generated at one or more previous iterations to produce an updated summary, and storing the updated summary via a data block of a distributed ledger.

Abstract: An example system includes a processor to prune a machine learning model based on an importance of neurons or weights. The processor is to further permute and pack remaining neurons or weights of the pruned machine learning model to reduce an amount of ciphertext computation under a selected constraint.

Abstract: A method, apparatus and computer program product for homomorphic inference on a decision tree (DT) model. In lieu of HE-based inferencing on the decision tree, the inferencing instead is performed on a neural network (NN), which acts as a surrogate. To this end, the neural network is trained to learn DT decision boundaries, preferably without using the original DT model data training points. During training, a random data set is applied to the DT, and expected outputs are recorded. This random data set and the expected outputs are then used to train the neural network such that the outputs of the neural network match the outputs expected from applying the original data set to the DT. Preferably, the neural network has low depth, just a few layers. HE-based inferencing on the decision tree is done using HE inferencing on the shallow neural network. The latter is computationally-efficient and is carried without the need for bootstrapping.

Abstract: An example operation may include one or more of generating, by a plurality of training participant clients, gradient calculations for machine learning model training, each of the training participant clients comprising a training dataset, converting, by a training aggregator coupled to the plurality of training participant clients, the gradient calculations to a plurality of transaction proposals, receiving, by one or more endorser nodes or peers of a blockchain network, the plurality of transaction proposals, executing, by each of the endorser nodes or peers, a verify gradient smart contract, and providing endorsements corresponding to the plurality of transaction proposals to the training aggregator.

Abstract: This disclosure provides a method, apparatus and computer program product to create a full homomorphic encryption (FHE)-friendly machine learning model. The approach herein leverages a knowledge distillation framework wherein the FHE-friendly (student) ML model closely mimics the predictions of a more complex (teacher) model, wherein the teacher model is one that, relative to the student model, is more complex and that is pre-trained on large datasets. In the approach herein, the distillation framework uses the more complex teacher model to facilitate training of the FHE-friendly model, but using synthetically-generated training data in lieu of the original datasets used to train the teacher.

Abstract: An example operation may include one or more of generating, by a training participant client comprising a training dataset, a plurality of transaction proposals that each correspond to a training iteration for machine learning model training related to stochastic gradient descent, the machine learning model training comprising a plurality of training iterations, the transaction proposals comprising a gradient calculation performed by the training participant client, a batch from the private dataset, a loss function, and an original model parameter, receiving, by one or more endorser nodes of peers of a blockchain network, the plurality of transaction proposals, and evaluating each transaction proposal.

Abstract: Privacy-preserving homomorphic inferencing utilizes batch processing on encrypted data records. Each data record has a private data portion of interest against which the inferencing is carried out. Batch processing is enabled with respect to a set of encrypted data records by techniques that ensure that each encrypted data record has its associated private data portion in a unique location relative to the other data records. The set of encrypted data records are then summed to generate a single encrypted data record against which the inferencing is done. In a first embodiment, the private data portions of interest are selectively and uniquely positioned at runtime (when the inferencing is being applied). In a second embodiment, the private data portions of interest are initially positioned with the data-at-rest, preferably in an off-line process; thereafter, at runtime individual encrypted data records are processed as necessary to adjust the private data portions to unique positions prior to batching.

Abstract: A method, apparatus and computer program product that provides multi-phase privacy-preserving inferencing in a high throughput data environment, e.g., to facilitate fraud prediction, detection and prevention. In one embodiment, two (2) machine learning models are used, a first model that is trained in the clear on first transaction data, and a second model that is trained in the clear but on the first transaction data, and user data. The first model is used to perform inferencing in the clear on the high throughput received data. In this manner, the first model provides a first level evaluation of whether a particular transaction might be fraudulent. If a transaction is flagged in this first phase, a second more secure inference is then carried out using the second model. The inferencing performed by the second model is done on homomorphically encrypted data. Thus, only those transactions marked by the first model are passed to the second model for secure evaluation.

Abstract: A technique for computationally-efficient privacy-preserving homomorphic inferencing against a decision tree. Inferencing is carried out by a server against encrypted data points provided by a client. Fully homomorphic computation is enabled with respect to the decision tree by intelligently configuring the tree and the real number-valued features that are applied to the tree. To that end, and to the extent the decision tree is unbalanced, the server first balances the tree. A cryptographic packing scheme is then applied to the balanced decision tree and, in particular, to one or more entries in at least one of: an encrypted feature set, and a threshold data set, that are to be used during the decision tree evaluation process. Upon receipt of an encrypted data point, homomorphic inferencing on the configured decision tree is performed using a highly-accurate approximation comparator, which implements a “soft” membership recursive computation on real numbers, all in an oblivious manner.

Abstract: An example operation includes one or more of computing, by a data owner node, updated gradients on a loss function based on a batch of private data and previous parameters of a machine learning model associated with a blockchain, encrypting, by the data owner node, update information, recording, by the data owner, the encrypted update information as a new transaction on the blockchain, and providing the update information for an audit.

Abstract: Aspects of the present disclosure describe techniques for detecting anomalous data in an encrypted data set. An example method generally includes receiving a data set of encrypted data points. A tree data structure having a number of levels is generated for the data set. Each level of the tree data structure generally corresponds to a feature of the encrypted plurality of features, and each node in the tree data structure at a given level represents a probability distribution of a likelihood that each data point is less than or greater than a split value determined for a given feature. An encrypted data point is received for analysis, and anomaly score is calculated based on a probability identified for each of the plurality of encrypted features. Based on determining that the calculated anomaly score exceeds a threshold value, the encrypted data point is identified as potentially anomalous.

Abstract: This disclosure provides a method, apparatus and computer program product to create a full homomorphic encryption (FHE)-friendly machine learning model. The approach herein leverages a knowledge distillation framework wherein the FHE-friendly (student) ML model closely mimics the predictions of a more complex (teacher) model, wherein the teacher model is one that, relative to the student model, is more complex and that is pre-trained on large datasets. In the approach herein, the distillation framework uses the more complex teacher model to facilitate training of the FHE-friendly model, but using synthetically-generated training data in lieu of the original datasets used to train the teacher.

Abstract: A technique for computationally-efficient privacy-preserving homomorphic inferencing against a decision tree. Inferencing is carried out by a server against encrypted data points provided by a client. Fully homomorphic computation is enabled with respect to the decision tree by intelligently configuring the tree and the real number-valued features that are applied to the tree. To that end, and to the extent the decision tree is unbalanced, the server first balances the tree. A cryptographic packing scheme is then applied to the balanced decision tree and, in particular, to one or more entries in at least one of: an encrypted feature set, and a threshold data set, that are to be used during the decision tree evaluation process. Upon receipt of an encrypted data point, homomorphic inferencing on the configured decision tree is performed using a highly-accurate approximation comparator, which implements a “soft” membership recursive computation on real numbers, all in an oblivious manner.

Abstract: A method, apparatus and computer program product for homomorphic inference on a decision tree (DT) model. In lieu of HE-based inferencing on the decision tree, the inferencing instead is performed on a neural network (NN), which acts as a surrogate. To this end, the neural network is trained to learn DT decision boundaries, preferably without using the original DT model data training points. During training, a random data set is applied to the DT, and expected outputs are recorded. This random data set and the expected outputs are then used to train the neural network such that the outputs of the neural network match the outputs expected from applying the original data set to the DT. Preferably, the neural network has low depth, just a few layers. HE-based inferencing on the decision tree is done using HE inferencing on the shallow neural network. The latter is computationally-efficient and is carried without the need for bootstrapping.

Abstract: An example operation includes one or more of computing, by a data owner node, updated gradients on a loss function based on a batch of private data and previous parameters of a machine learning model associated with a blockchain, encrypting, by the data owner node, update information, recording, by the data owner, the encrypted update information as a new transaction on the blockchain, and providing the update information for an audit.

Abstract: In an approach for training machine-learning models using encrypted data, a processor receives a set of encrypted data from a client computing device. A processor trains a machine-learning model using a boosting algorithm. A processor performs a first classification on the set of encrypted data using the machine-learning model. A processor sends a first set of encrypted results of the first classification to the client computing device. A processor receives a first set of boosting updates from the client computing device. A processor applies the first set of boosting updates to the machine-learning model.