This work package focuses on problems of statistical estimation, hypothesis testing and decision-making in distributed systems, bearing in mind additional constraints such as communication and privacy.
Their work to date includes a new method in how to securely and reliably store information in a distributed setting when there’s a malicious hacker who can only see and corrupt (an unknown) part of the system. Optimal Information Security Against Limited-View Adversaries: The Benefits of Causality and Feedback – University of Bristol and Codes for Adversaries: Between Worst-Case and Average-Case Jamming.
In addition, researchers have conducted an exploration into how to split large datasets into smaller, less similar batches to improve machine learning performance; intentionally mixing dissimilar data in each batch to help models learn more general patterns and avoid overfitting to repetitive or redundant data Dissimilar Batch Decompositions of Random Datasets | Sankhya A
