Abstract: Disclosed herein are systems and methods for securely executing a machine learning model distributed over at least one client device and at least one server. In one aspect, a method comprises determining whether a first operation performed by the model can be reduced to one or more operations of a specific type compatible with a first encryption scheme. When the first operation can be so reduced, data associated with the first operation is encrypted using the first encryption scheme. When the first operation cannot be so reduced, the data is encrypted using a second encryption scheme. The encrypted data is then transmitted to the at least one server configured to apply the first operation.

Abstract: A system receives a user query requesting a testable hypothesis about a scientific topic. The system classifies the user query into a first theoretical framework of a plurality of theoretical frameworks each comprising of terms and principles related to a particular scientific topic. The system generates the testable hypothesis by a first machine learning (ML) model that is configured to: receive as inputs: the user query, the first theoretical framework, and information from a graph document database comprising data associated with scientific documents, generate, as an output, the testable hypothesis that can be evaluated using the first theoretical framework and that does not reiterate a hypothesis or findings from the scientific documents in the graph document database. The system outputs the testable hypothesis via a user interface in response to the user query.

Abstract: A system performs open set detection on each respective video frame from a plurality of video frames, wherein the open set detection comprises: generating one or more bounding boxes around brand media in the respective video frame, determining embedding vectors for each of the one or more bounding boxes; and comparing the embedding vectors with known embedding vectors of one or more brand reference images. The system segments the brand media within the one or more bounding boxes by determining an exact region in which a brand logo is occupied within the one or more bounding boxes using a semantic segmentation model. The system computes one or more parameters associated with per-brand, per-appearance statistics of the brand logo in the video based on the comparing.

Abstract: Systems and methods provide a multidimensional data structure to tailor data storage and corresponding user output. Tailored graph data structures reduce the level of hallucination in composing an answer to a user using a large language model. In response to a user query of documents, the user query is translated to a vector. From the vector, the most relevant pieces of documents that are as similar as possible in semantics to the question of a person can be determined and communicated to the user. In one aspect, using the vector base (vector query applied to the documents), a graph can be generated in the form of nodes and connections between data in the documents.

Abstract: A rootkit detection system and method analyzes memory dumps to determine connections between intercepted system driver operations requested by unknown files and changes in system memory before and after those operations. Memory dump differences and I/O buffers are analyzed with machine learning models to identify clustered features associated with rootkits.

Abstract: Disclosed herein are systems and methods for updating a graphical user interface displaying content related to a topic based on user performance. A method may include: receiving, via a user interface (UI), a user selection of a topic; generating, using a machine learning algorithm, a course from reference materials describing a plurality of sub-topics associated with the topic, wherein the machine learning algorithm includes, in the course, a plurality of assessments that test comprehension of the plurality of sub-topics; outputting the course on the UI; monitoring user interaction with a first subset of the assessments within the course on the UI; in response to determining, based on the monitoring, that the user interaction does not meet a comprehension criteria, modifying, using the machine learning algorithm, a first subset of the sub-topics corresponding to the first subset of the assessments; and outputting the modified course on the UI.

Abstract: Disclosed herein are systems and methods for generating custom large language model (LLM) embeddings. The method includes: receiving an input query to a LLM service from a user; determining whether custom embeddings for the input query are stored in a cache table, based on a determination that the custom embeddings for the input query are stored in the cache table, sending the custom embeddings from the cache table to the LLM service; and based on a determination that the custom embeddings for the input query are not stored in the cache table, executing a prepared embeddings cache model (ECM) to analyze the input query, generating custom embeddings for the input query based on results from prepared ECM, and transmitting the custom embeddings to the LLM service.

Abstract: Disclosed herein are systems and methods for training an embeddings cache model (ECM). The method includes generating a main embedding by a main embedding model based on a received input query. The method also includes searching for the input query or the generated main embedding in a cache table. Based on a determination that the input query is not found in the cache table and the main embedding is not found in the cache table within the embedding similarity threshold, the method further includes: based on a determination that the number of entries in the cache table is equal to or greater than the predetermined threshold for training the ECM, storing the input query and the main embedding in the cache table, and training or fine-tuning the ECM based on the input query and the main embedding, and transmitting the main embedding to the LLM service.

Abstract: Disclosed herein are systems and methods for training an embedding transformation model to predict a main embedding. The method includes receiving an input query to a LLM service. The method also includes determining if a main embedding of the query is provided by a main embedding model. The method also includes generating a secondary embedding of the query by a pre-trained fallback embedding model. The method further includes based on a determination that the main embedding of the input query is provided by the main embedding model, preparing a embedding transformation model using the main and secondary embeddings to predict the main embedding based on the secondary embedding.

Abstract: Disclosed herein are systems and methods for generating LLM embeddings. The method includes receiving an input query to a LLM service. The method also includes receiving an input query to a LLM service; determining if a main embedding of the query is provided by a main embedding model; based on a determination that the main embedding of the input query is not provided by the main embedding model, generating a secondary embedding of the query by a pre-trained fallback embedding model, and applying an embedding transformation model to the secondary embedding to predict the main embedding and transmitting the predicted main embedding to the LLM service.

Abstract: Disclosed herein are systems and methods for providing LLM embeddings. The method also includes receiving a query to a LLM service. The method also includes determining if a first embedding of the input query is provided by a main embedding model. The method further includes, based on a determination that the first embedding is provided by the main embedding model, identifying an embeddings cluster similar to the input query, and computing a main centroid embedding for the embeddings cluster based on the first embedding. The method further includes based on a determination that the first embedding is not provided by the main embedding model, identifying an embeddings cluster similar to the input query, and transmitting the main centroid embedding of the embeddings cluster to the LLM service.

Abstract: Disclosed herein are systems and method for repurposing a machine learning model. An exemplary method includes: receiving a first training dataset; determining an input portion and an output portion in an entry of the first training dataset; comparing the first training dataset to a second training dataset used to train a machine learning model, wherein the comparing includes determining a similarity score between the input portion and the output portion of the first training dataset and an input portion and an output portion of the second training dataset; in response to determining that the similarity score is greater than a threshold similarity score, re-training the machine learning model using the first training dataset; and executing the re-trained machine learning model on an input value to generate an output value corresponding to the first training dataset.

Abstract: Disclosed herein are systems and methods for providing LLM embeddings. The method also includes receiving a query to a LLM service. The method also includes determining if a first embedding of the input query is provided by a main embedding model. The method further includes, based on a determination that the first embedding is provided by the main embedding model, identifying an embeddings cluster similar to the input query, and computing a main centroid embedding for the embeddings cluster based on the first embedding. The method further includes based on a determination that the first embedding is not provided by the main embedding model, identifying an embeddings cluster similar to the input query, and transmitting the main centroid embedding of the embeddings cluster to the LLM service.

Abstract: A system determines whether a first operation performed by an MLM is compatible with one of a first encryption scheme and a second encryption scheme, wherein the MLM is distributed over at least one client device and at least one server. In response to determining that the first operation is compatible with the first encryption scheme, the system: encrypts data associated with the first operation using the first encryption scheme; and transmits the data encrypted by the first encryption scheme to the at least one server configured to apply the first operation. In response to determining that the first operation is incompatible with the first encryption scheme, the system: encrypts the data associated with the first operation using the second encryption scheme; and transmits the data encrypted by the second encryption scheme to the at least one server configured to apply the first operation.

Abstract: Aspects of the present disclosure include a method for verifying live user presence in an online session, comprising receiving first and second video streams capturing a user during the session from a first position and a different second position, respectively. The method further comprises initiating a challenge by providing, for presentation on a different second display of a second device, a code, and providing, for presentation on a first display, an instruction to the user to respond to the challenge using the second device. The method further comprises detecting, based on at least one of the video streams, physical actions of the user during the challenge, receiving, from the second device, user input events representing a user response to the challenge, and determining, based on at least one of the physical actions or the user input events, whether the user successfully completed the challenge.

Abstract: A method and system is disclosed for generating video content where speech audio is accurately synchronized with facial movements. The method begins by converting input text into speech audio using a text-to-speech (TTS) system. An original video is then processed to identify and extract the mouth area of the human face. The method generates a plurality of frames of the mouth area corresponding to segments of the speech audio using a trained generative neural network. These frames are lip-synced with the corresponding audio segments. The extracted mouth area frames are overlaid onto the original video, followed by performing facial reconstruction using a trained facial reconstruction model to correct distortions and achieve a natural and expressive appearance around the mouth area. Finally, the generated speech audio is combined with the modified video to produce a final video with synchronized lip movements and phonetic sounds.

Abstract: Aspects of the present disclosure include a method for verifying live user presence in an online session, comprising obtaining at least one video stream of a user captured during at least one previous presentation of at least one user interface (UI) on a display, obtaining input events representing previous user interactions with the UI, generating a behavioral digital fingerprint profile based on the video stream and the input events, initiating an online exam during the online session by providing, for presentation on the display, at least one additional UI, receiving at least one additional video stream of the user during the exam, receiving additional input events representing user interactions with the additional UI, determining exam-specific behavioral features based on the additional video stream and the additional input events, and verifying whether the user is taking the exam based on at least one of the features or the profile.

Abstract: Disclosed herein are systems and method for a dynamic knowledge graph (KG) augmentation of an LLM. The method includes obtaining a LLM chat history between a user and the LLM enhanced with a KG database. The LLM chat history includes user queries, relevant knowledge data retrieved from the KG database, and LLM answers for each user query. The LLM answers are based at least in part on relevant knowledge data retrieved from the KG database. The method also includes applying a KG update MLM agent trained to: analyze the LLM chat history, KG database and the graph schema and to identify: (i) a missing knowledge data from the KG database, and/or (ii) a missing relationship between knowledge data nodes in the KG database. The method further includes after identifying missing knowledge data and/or the missing relationship, updating the KG database with new knowledge data and/or new relationship between nodes.

Abstract: Systems and methods provide a multidimensional data structure to tailor data storage and corresponding user output. Tailored graph data structures reduce the level of hallucination in composing an answer to a user using a large language model. In response to a user query of documents, the user query is translated to a vector. From the vector, the most relevant pieces of documents that are as similar as possible in semantics to the question of a person can be determined and communicated to the user. In one aspect, using the vector base (vector query applied to the documents), a graph can be generated in the form of nodes and connections between data in the documents.

Abstract: Aspects of the present disclosure include a method for verifying live user presence in an online session. The method comprises providing, for presentation, a time-varying pattern and an instruction to a user to perform one or more requested user actions in synchrony with the pattern. The method further comprises receiving one or more data streams of the user during the presentation, extracting from the data streams feature information indicative of one or more detected user actions in the data streams, and generating one or more event signals based on the feature information. The method further comprises determining, based on the event signals and the pattern, a first measurement and a second measurement indicative of spatial correspondence and temporal alignment, respectively, between the detected user actions and the pattern, and verifying whether the user is a live person based on the first and second measurements.