Abstract: In some implementations, a device may cluster a set of jailbreak attack templates to form an attack style library according to attack styles. The device may select a subset of clusters based on a red-teaming goal to create a combination of different attack styles. The device may generate diverse seed prompts based on the selected subset of clusters. The device may apply a branching phase to each seed prompt to generate a plurality of prompt samples. The device may apply a similarity filter algorithm to remove redundant prompt samples from the prompt samples. The device may evaluate remaining prompt samples to determine successful prompt samples that meet a jailbreak success threshold. The device may output the successful prompt samples.

Abstract: A computer-implemented method, a computer program product, and a computer system for recommending model contributions based on federated learning lineage. The computer system retrieves information of model checkpoints. The computer system trains data analytic models for monitoring activities of training rounds in a federated learning system, based on the information of the model checkpoints. The computer system sends to a user summary statistics of the model checkpoints. The computer system receives from the user natural language instructions of modifying a federated learning plan for future training rounds in the federated learning system. The computer system translates the natural language instructions into updates for the federated learning system. The computer system forwards the updates to the federated learning system.

Abstract: A method of generating machine learning predictions by an efficiently updatable ensemble of machine learning models includes identifying, in response to an inference request, one or more machine learning models of the ensemble that are available for generating a prediction. An aggregated prediction is generated in response to the inference request. The aggregated prediction aggregates individual predictions generated by the one or more machine leaning models of the ensemble identified as available to generate a prediction. Responsive to determining that less than all of the ensemble of machine learning models are available, a performance guarantee based on the individual predictions is generated. The aggregated prediction is output in response to the performance guarantee satisfying a predetermined threshold.

Abstract: Systems, devices, computer program products and/or computer-implemented methods of use provided herein relate to federated training and inferencing. A system can comprise a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory, wherein the computer executable components can comprise a modeling component that trains an inferential model using data from a plurality of parties and comprising horizontally partitioned data and vertically partitioned data, wherein the modeling component employs a random decision tree comprising the data to train the inferential model, and an inference component that responds to a query, employing the inferential model, by generating an inference, wherein first party private data, of the data, originating from a first passive party of the plurality of parties, is not directly shared with other passive parties of the plurality of parties to generate the inference.

Abstract: Embodiments for providing expert-in-the-loop training of machine learning models in a computing environment by a processor. A performance of a machine learning model may be learned. Feedback for the machine learning model may be received based on learning the performance the machine learning model, where the feedback includes domain knowledge provided by a domain expert. The machine learning model may be trained or updated based the feedback of the performance of the machine learning model.

Abstract: The present disclosure relates to a method comprising a training system iteratively training a machine learning algorithm using current training data. The current training data comprises a local dataset of a current task and a replay dataset and may be updated for a next iteration as follows. A training dataset may be received. If the training dataset is not a shared dataset and its task is different from the current task: information representing the local dataset may be shared with other training systems, the local dataset may be added to the replay dataset, and the received training dataset may be used as the local dataset for a next iteration. In case the task is the current task: the received training dataset may be added to the local dataset. If the training dataset is a shared dataset, the received training dataset may be added to the replay dataset.

Abstract: A computer-implemented method, a computer program product, and a computer system for assessing fairness of a deep generative model. A computer system receives a user defined fairness criterion for the deep generative model. A computer system probes the deep generative model to produce samples for a target output. A computer system evaluates the samples for the fairness of the deep generative model, according to the user defined fairness criterion. A computer system produces a set of recommendations for modifying the deep generative model to meet the user defined fairness criterion, in response to determining that the deep generative model does not meet the user defined fairness criterion. In response to determining that the deep generative model is to be modified, a computer system applies at least one subset of the recommendations to the deep generative model. A computer system updates the deep generative model.

Abstract: An exemplary system comprises a memory that stores and a processor that executes computer executable components stored in the memory, wherein the computer executable components comprise an obtaining component that intercepts a semantic source from being submitted to a retrieval augmented generation (RAG) architecture, and a transforming component that transforms the semantic source into a transformed source by identifying and converting prompt-misleading text of the semantic source into prompt-non-misleading text. In one or more embodiments, the semantic source is a semantic query having been submitted to the RAG architecture and/or a retrieved source having been retrieved by the RAG architecture in a process of providing a prompt. In one or more embodiments, the prompt-misleading text originated in connection with an origination of the semantic source and/or was caused by an adversarial attack corresponding to the semantic source.

Abstract: A method, computer program, and computer system are provided for predicting and assessing risks on websites. Data corresponding to historical interactions of a user with one or more websites is accessed. A simulation of actions of the user is generated based on the accessed data, and actions of the user are simulated on a pre-defined target website based on the generated simulation of the actions of the user. Risks on the target website are identified based on simulating the actions of the user. The website is updated to mitigate the identified risks.

Abstract: Secure noise addition in floating-point numbers is provided. It is determined whether digits of a mantissa of a summed floating-point number include a set of trailing zeros at an end of the mantissa of the summed floating-point number. In response to determining that the digits of the mantissa of the summed floating-point number include the set of trailing zeros at the end of the mantissa of the summed floating-point number, the set of trailing zeros at the end of the mantissa of the summed floating-point number is replaced with a set of digits selected from a group of random digits to form an output floating-point number that is free from traces of a sensitive non-integer input value satisfying differential privacy guarantee of data security immune from floating-point attack.

Abstract: Generating performance metrics and recommendations to improve an unlearned model includes executing an unlearning algorithm to expunge the influence on a machine learning model of a selected sample of the machine learning model's training dataset. Executing the unlearning algorithm creates an unlearned model. Performance metrics are generated by a metrics generator for the unlearned model and the machine learning model. Based on the performance metrics, an unlearning analysis is generated by a comparator comparing the performances of the unlearned model and machine learning model. A recommender, based on the unlearning analysis, generates a recommendation recommending a revision to the unlearned model in response to detecting a deviation of more than a predetermined threshold of one or more of the performance metrics of the unlearned model from one or more of the performance metrics of the machine learning model. An evaluator generates an unlearning evaluation of the unlearned model.

Abstract: An embodiment trains, using a database of tasks, a classifier model to classify an input task into a task category. An embodiment generates a plurality of prompts. An embodiment applies a first prompt in the plurality of prompts to a trained model, the trained model producing a first model output in response to the first prompt. An embodiment classifies, using the trained classifier model, the first model output into a first task category. An embodiment determines that the first task category is an undesired task category. An embodiment adjusts, responsive to determining the first task category is the undesired task category, the trained model, the adjusting altering a capability of the trained model to perform a task in the first task category.

Abstract: A method for additive manufacturing includes identifying a discrepancy between a three-dimensional model and an object model. The three-dimensional model is a model of a three-dimensional object that is being constructed by an additive manufacturing process, and the three-dimensional object is being constructed based on the object model. The method further includes determining a reconfiguration recommendation based on the identified discrepancy. The method further includes reconfiguring the additive manufacturing process based on the reconfiguration recommendation.

Abstract: One or more systems, devices, computer program products and/or computer-implemented methods of use provided herein relate to facilitating a process to compensate a service being provided over a network connection. A system can comprise a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory, wherein the computer executable components can comprise a determination component that determines a network connection between a server and a client node, and a predictive component that predicts, employing machine learning, a graphical representation update to a service provided by the server over the network connection. The predictive component can generate the prediction in response to a decrease in bandwidth and/or an increase in latency of a network connection. A training component can train a machine learning model employed by the predictive component based on historical data of the service provided by the server.

Abstract: A system, computer program product, and method are presented for administering examinations with adversarial hardening of queries against automated responses. The method include receiving an original query electronically. A response to the original query is to be submitted electronically by a human. The method also includes modifying the original query, thereby generating a modified query. The modified query is configured to be comprehensible by the human, and not properly responded to through electronic means without human support.

Abstract: An embodiment causes generating, by a trained model, a training prompt response to a training prompt in a set of training prompts. An embodiment trains, using the training prompt and the training prompt response, an attribution model, the training resulting in a trained attribution model. An embodiment attributes, using the trained attribution model and a first prompt response generated by a fine-tuned model in response to a prompt, the fine-tuned model to a foundation model.

Abstract: Adversarial attack detection operations may be applied on one or more deep generative models for defending deep generative models from adversarial attacks. The adversarial attack may be detected on the one or more deep generative models based on the one or more of a plurality of adversarial attack detection operations. The one or more deep generative models may be sanitized based on the adversarial attack.

Abstract: Techniques of facilitating deep learning model rescaling by computing devices. In one example, a system can comprise a processor that executes computer executable components stored in memory. The computer executable components can comprise: a rescaling component; and a forecasting component. The rescaling component can determine a scaling ratio that maps low mesh resolution predictive data output by a partial differential equation (PDE)-based model for a sub-domain to high-resolution observational or ground-truth data for a domain comprising the sub-domain. The forecasting component can generate high mesh resolution predictive data for the domain with a machine-learning model using input data of the PDE-based model and the scaling ratio.

Abstract: A system includes an agent engine, an encoder, a general-purpose solver engine, and an orchestrator. The orchestrator is configured to receive a first problem instance corresponding to a learned policy that is based on auto reinforcement learning, and provide the first problem instance to the general-purpose solver engine, which is configured to execute based on the first problem instance to determine a solver state. The orchestrator is configured to extract, from the general-purpose solver engine, the solver state, and to provide the solver state to the encoder. The encoder is configured to query the agent engine for a best action according to the learned policy and an encoded solver state. The agent engine is configured to determine the best action according to the learned policy and the encoded solver state. The orchestrator is configured to receive the best action, and direct the general-purpose solver to implement the best action.

Abstract: Methods, computer program products and/or systems are provided that perform the following operations: obtaining a collection of objects; constructing a weighted graph based on the collection of objections, wherein the weighted graph preserves neighborhood semantics of objects of the collection of objects; generating partitions of nodes in the weighted graph of a fixed maximum size utilizing combinatorial partitioning; generating a vector for each node based on the partitions of nodes in the weighted graph; determining vector representations for objects in this collection and eventually applying this vector representation of the objects to gain efficiency (e.g., in terms of computation time and memory requirements) for use in downstream tasks such as recommendation.