Archive
Recent publications, grants, awards, and group updates.
Accepted to ICML 2026, this work studies how unlearning transfers across languages and how unlearned information can be recovered through activation steering.
Presented at ICLR 2026, this work gives new theory for disagreement discrepancy surrogates under distribution shift and introduces a provably consistent alternative.
Accepted at the 18th NASA Formal Methods Symposium, this work explores applying automated reasoning to the problem of verifying binaries.
This NeurIPS 2025 paper studies adaptive data analysis in settings where the underlying dataset grows over time.
AdaptDel was presented at NeurIPS 2025 and develops certified robustness guarantees for sequence models under edit-distance perturbations.
V-ORAM appeared at the 34th USENIX Security Symposium and proposes an adaptive ORAM framework for changing workloads.
THEMIS was presented at the 34th USENIX Security Symposium in Seattle and studies practical protection mechanisms for deployed on-device deep learning models.
Presented at FUZZING 2025, co-located with ISSTA 2025, Trailblazer introduces an end-to-end workflow for testing web APIs even when no machine-readable specification is available.
The paper, coauthored with collaborators at George Washington University and Barnard College, was accepted to the 46th IEEE Symposium on Security and Privacy.