Abstract: One example method includes receiving a text string that includes multiple words, tokenizing the text string to create a tokenized text string, substituting each token in the tokenized text string with a mask token to create a masked text string, performing an inference process on the masked text string to obtain a respective probability for each token, determining a respective suspicion level for each probability, modulating the suspicion levels to obtain a respective weighted suspicion score for each token, and comparing each of the weighted suspicion scores with a threshold to determine whether any one or more of the words indicate that the text string includes an attack prompt.

Abstract: One example method includes receiving a data file at a large language model (LLM). Arbitrary tags that include labels that are attachable to the data file and prompts are also received. The prompts are paired with the arbitrary tags to form arbitrary tag-prompt pairs and include information that is used by the LLM to find the paired arbitrary tag. The LLM determines a selected subset of the arbitrary tags that apply to the data file. A trust module receives the selected subset of the arbitrary tags that apply to the data file and data access policies that specify access rules for the data file. A conditional access decision is determined that specifies whether access should be given to the data file.

Abstract: One example method includes, for each federated learning simulation, defining a machine learning model that is used in the federated learning simulation. The machine learning model has associated variables and is implemented at edge nodes and a central node of the federated learning simulation. A first variable list is defined that specifies associated variables that are to be optimized at the edge nodes of the federated learning simulation. A second variable list is defined that specifies associated variables that are to be provided by the edge nodes to the central node of the federated learning simulation. The associated variables included in the first variable list are optimized at the edge nodes of federated learning simulation. The associated variables that are included in the second variable list and that are provided by the edge nodes of the federated learning simulation are aggregated by the central node of the federated learning simulation.

Abstract: One example method includes a machine-learning (ML) model receiving a first input that includes images that have been extracted from a web page and a second input that includes alt-texts that have been extracted from the web page. The alt-texts describe the images. The ML model converts the images into a first embedding representation and converts the alt-texts into a second embedding representation. Based on the first and second embedding representations, a similarity score between the images and the alt-texts is calculated. The similarity score specifies how accurately each of the alt-texts describe the images. The one of the alt-texts having the highest similarity score is then selected.

Abstract: One example method includes receiving input from a user, the input including reference information, and a document corpus that comprises a group of documents, performing a byte pair encoding (BPE) process, and/or preprocessing, on the documents in the document corpus, so as to generate a respective TDF-IDF (term frequency-inverse document frequency) vector for each of the documents in the document corpus, comparing each of the TDF-IDF vectors to the reference information, and based on the comparing, ranking the documents according to their respective relevance to the reference information.

Abstract: One example method includes accessing an adversarial image, processing the adversarial image by applying a smoothing defense to the adversarial image, and classifying the processed adversarial image. The smoothing defense, which may be an unsupervised process, may include a noising process, such as a gaussian noising process, and an edge-preserving process. The smoothing defense can be implemented against basic iterative method (BIM) attacks, as well as fast gradient sign method (FGSM) attacks.

Abstract: Saliency heatmap generation is disclosed. A mask that corresponds to an image is iteratively passed through a model. At each iteration, the output mask is optimized via gradient descent to gradually turn off one or more pixels. The output mask is iteratively passed in order to minimize contributions of pixels and to maximize a class prediction. After generating a final output mask, a saliency heatmap is generated.

Abstract: A method includes assembling an explainable artificial intelligence committee comprising two or more explainable artificial intelligence techniques, performing the explainable artificial intelligence techniques on results generated by a machine learning model, as a result of the performing, obtaining respective explanations, generated by each of the explainable artificial intelligence techniques, for the results generated by the machine learning model, and determining that one of the explanations was compromised by an attacker.

Abstract: One method includes stochastically selecting, by a central node, a subset of edge nodes from a group of edge nodes that collectively defines a federation, querying, by the central node, the edge nodes of the subset for updates to a global model maintained by the central node, receiving, by the central node from the edge nodes of the subset, respective updates to one or more layers of the global model, and updating, by the central node, the global model, using the updates received from the edge nodes of the subset.

Abstract: Techniques are provided for evaluation of machine learning models using agreement scores. One method comprises obtaining two or more of: (i) a first set of quantitative features characterizing model parameters of a machine learning model; (ii) a second set of quantitative features characterizing a training process used to train the machine learning model; and (iii) a third set of quantitative features characterizing a training dataset used to train the machine learning model; generating a score based on an aggregation of at least portions of the two or more of the first set, the second set and the third set, wherein the score is based on an agreement of the machine learning with designated characteristics; and initiating an automated action based on the score. The automated action may comprise updating the machine learning model; generating a notification in connection with an audit; and/or selecting a machine learning model for deployment.

Abstract: One example method includes initiating an audit of a machine learning model, providing input data to the machine learning model as part of the audit, while the audit is running, receiving information regarding operation of the machine learning model, wherein the information comprises a computational resource footprint, analyzing the computational resource footprint, and determining, based on the analyzing, that the computational resource footprint is characteristic of an adversarial attack on the machine learning model.