Abstract: Identifying malicious clients in federated learning is disclosed while enhancing privacy. The clients are clustered such that cluster updates in the federated learning are generated. When a suspect cluster is identified, clients in the suspect clusters are labeled as suspect and clients in clusters that are not suspect are labeled as fair. The clients are reclustered and the clusters and clients are relabeled without changing the labels of clients that were previously deemed fair. After one or more iterations, the malicious clients are identified, and corrective actions can be performed.

Abstract: Model selection is disclosed. Features used as inputs to models are scored in terms of importance and health. The importance and health scores are combined in order to generate a model score for each model. The model with a score above a threshold score is selected and deployed.

Abstract: One example method includes constructing a machine learning model which, when completed, is operable to screen candidates, from a group of candidates, to define a candidate pool that has specified characteristics. The constructing includes: broadcasting, from a central node to edges of a federation, an indication that construction of a random forest, of the machine learning model, has started; performing a federated feature categorization, by the central node based on information received from the edges, of a feature to be included in respective decision trees of the edges; based on the categorizing, broadcasting a feature category to the edges; performing, by the central node using respective purity information received from the edges, a federated purity calculation; and based on the federated purity calculation, broadcasting, by the central node to the edges, a winning feature split for the feature.

Abstract: Unsupervised event detection is disclosed. Reconstruction data resulting from processing input samples with a machine learning model is clustered. By labeling one or more samples of a cluster, all of the samples in the same cluster can be labeled the same. During inference, any input sample generating a similar reconstructed sample can be given the label previously applied to the cluster.

Abstract: One example method includes receiving, by a central node, respective data and statistics from each edge node in a group of edge nodes, performing, by the central node, a similarity-based clustering of the edge nodes so that different clusters of edge nodes are defined, sampling, by the central node, edge nodes from each of the clusters to perform a quantization selection method, receiving, by the central node, from the sampled edge nodes, a respective indication of a best-performing quantization method, and electing, by the central node, a quantization method to be used by all the edge nodes.

Abstract: One example method includes receiving at a client node of a federation a global machine-learning model that is to be trained by the client node using a training dataset that is local to the client node. In response to receiving the global machine-learning model, determining at the client node if the global machine-learning model is trending toward an overfitted state using a validation dataset. The overfitted state indicates that the global machine-learning model has not been received from a server that is part of the federation because of a client isolation attack. In response to determining that the global machine-learning model is trending towards the overfitting state, causing the client node to leave the federation. In response to determining that the global machine-learning model is not trending towards the overfitted state, training the global machine-learning model using the training dataset to thereby update the global machine-learning model.

Abstract: One example method includes running a workload through a trained open-set classification model, recovering, as a result of the running, a class and an open-setness score corresponding to the workload, determining, based on the class and the open-setness score, whether the workload is new, and when the workload is determined to be new, starting a new cluster that includes the workload. A response time predictor model may be used to predict a response time associated with the new workload.

Abstract: One example method includes receiving a rule-set, including a combination of rules, that was determined to occur in a set of ground truth documents, applying the rule-set to a new document that was not included in the set of ground truth documents, determining whether or not a rule in the rule-set succeeded or failed when applied to a word in the new document, and when the rule is determined to have failed, identifying the failed rule, identifying a confidence level in the determination that the rule failed, and when the confidence level is below a threshold confidence level, identifying the word, to which the failed rule was applied, as a candidate for verification by a human.

Abstract: One example method includes registering, by a customer, with a service provider, receiving, by the customer from the service provider, a global machine learning model, running, by the customer, the global machine learning model as a local machine learning model, collecting, by the customer, unlabeled data generated by edge devices operating in a customer domain, checking, by the customer, to determine if the customer domain has changed, and when it is determined that the customer domain has changed, performing, by the customer, a model adaptation process on the local machine learning model, and transmitting to the service provider, by the customer, gradients that comprise customer implemented changes to the local machine learning model.

Abstract: One example method includes transmitting, by a central node to each edge node in a group of edge nodes, a quantization level, receiving, by the central node from each of the edge nodes, a respective gradient vector, wherein each gradient vector has been quantized according to the quantization level, re-quantizing, by the central node, the gradient vectors that have been received from the edge nodes, wherein the gradient vectors are re-quantized by the central node to a lower quantization level than the quantization level, validating, by the central node, the quantization level and the lower quantization level, and based on an outcome of the validating, automatically adjusting the quantization level.

Abstract: One example method includes receiving a random forest classifier model that comprises a group of decision trees, wherein the random forest classifier model is created using a vertical federated framework, providing new observations, not included in a set of training observations, to a trained random forest classifier model, wherein the random forest classifier model is trained in the vertical federated framework, and wherein the training is performed using the set of training observations as input to the random forest classifier model, and generating, by the trained random forest classifier model, one or more diversity scores pertaining to the new observations.

Abstract: Training examples are created from telemetry data, in which each training example engineered features derived from the telemetry data, storage system characteristics about the storage system that processed the workload associated with the telemetry data, and the response time of the storage system while processing the workload. The training examples are provided to an unsupervised learning process which assigns the training examples to clusters. Training examples of each cluster are used to train/test a separate supervised learning process for the cluster, to cause each supervised learning process to learn a regression between independent variables (system characteristics and workload features) and a dependent variable (storage system response time). To determine a response time of a proposed storage system, the proposed workload is used to select one of the clusters, and then the trained learning process for the selected cluster is used to determine the response time of the proposed storage system.

Abstract: Techniques described herein relate to a method for predicting field values of documents. The method may include identifying a field prediction model generation request; obtaining, training documents from a document manager; selecting a first training document; making a first determination that the first training document is a text-based document; performing text-based data extraction to identify first words and first boxes included in the first training document; identifying first keywords and first candidate words included in the first training document based on the first words and the first boxes; and generating a first annotated training document using the first keywords and the first candidate words, wherein the first annotated training document comprises color-based representation masks for the first keywords, the first candidate words, and first general words included in the first training document.

Abstract: Object driven event detection is disclosed for nodes in an environment. Video frames of interest are identified from the video streams of cameras in the environment. The video frames of interest are input, along with node positions for nodes in the area of coverage of the cameras, into a detection module. The output of the detection model, combined with the output of an event model, are used by a decision pipeline to make decisions and perform actions in the environment.

Abstract: Techniques are disclosed for dynamic edge-weighted quantization. For example, a system can include at least one processing device including a processor coupled to a memory, the at least one processing device being configured to implement the following steps: selecting edge nodes for sampling based on an edge node sampling algorithm configured to use a specified number of edge nodes to be sampled; causing the selected edge nodes to execute a quantization selection procedure; receiving, from the selected edge nodes, identifications of a quantization procedure based on the quantization selection procedure; and selecting a quantization procedure for each edge node, based on the identifications of the quantization procedures for the selected edge nodes.

Abstract: One example method includes running an edge node sampling algorithm using a parameter ‘s’ that specifies a number of edge nodes to be sampled, using historical statistics from the edge nodes, calculating a composite time for each of the edge nodes, and the composite time comprises a sum of a federated learning time and an execution time of a quantization selection procedure, identifying an outlier boundary, defining a cutoff threshold based on the outlier boundary, and selecting, for sampling, the edge nodes that are at or below the cutoff threshold.

Abstract: One example method includes, for each document in a group of annotated documents, extracting a set of words from the annotated document, and each of the words is positioned in a respective field of the annotated document. The method further includes using an aggregation function to determine, for one of the fields, a similarity of each one of the annotated documents to all of the other annotated documents, creating a document layout graph with nodes that each correspond to a respective annotated document, and each node is connected to all other nodes for which a similarity threshold for the one field has been met, and running an algorithm on the document layout graph to identify a clique of the annotated documents, and each annotated document in the clique has a similar layout to respective layouts of the other annotated documents in the clique.

Abstract: Real-time event detection is performed on nodes in an environment using position data that is not available to a node in real time but is delayed. A node performs real time event detection by predicting a position of the node based at least in part on delayed position data. The delayed position data is aligned to other sensor data. Aligning the position data may include predicting a position based on dead reckoning and/or a machine learning model. One or more collections of data, each collection including sensor data and predicted position data, is input to a model that performs event detection.

Abstract: Model assessment is disclosed. When a model operates, tuples are transmitted to a central node. The central node can process the tuples received from multiple nodes to generate an efficiency score for the model. The efficiency score reflects how the inference of the model correlates to operator actions. Models whose assessment is below a threshold score may be retrained at least for certain classes.

Abstract: One example method includes training an event predictor, of a machine learning model, wherein the training includes training an event predictor of the machine learning model, and training open set trajectory classifier, of the machine learning model. After the event predictor is trained, the event predictor is operable to receive an unknown trajectory class and predict an event class for the unknown trajectory class. Further, after the open set trajectory classifier is trained, the open set trajectory classifier is operable to receive an unknown trajectory and classify the unknown trajectory with a predicted trajectory classification. Finally, during training, the output of the open-set trajectory classifier may be used as an input to the event predictor.