Project Grant 2212458

This three-year $300,000 project grant from the National Science Foundation's Computer and Information Science and Engineering program aims to advance understanding of robustness in machine learning models. Specifically, the University of Maryland, College Park will research conditions under which adversarial attacks on deep networks can be detected and original data reconstructed. It will also study fundamental limits of robustness guarantees against poisoning attacks, especially with a constant fraction of training data poisoned. Non-linear predictors exploiting sparsity and local stability will be analyzed for provable guarantees of robustness. The role of symmetry as a form of parsimony increasing adversarial robustness will also be examined. The award supports the NSF's mission to advance computing and communication foundations research and development.