Forensic Accountability in Heterogeneous Federated Learning
Federated learning can protect raw data locality while still remaining vulnerable to malicious client behavior. Existing Byzantine-robust aggregation methods primarily focus on suppressing or rejecting suspicious updates, but they do not inherently answer two forensic questions: which client was responsible for malicious behavior, and when did that behavior begin?
Research Direction
Develop lightweight forensic mechanisms that operate on server-observed client-update histories and preserve client identity across communication rounds. The goal is to distinguish persistent malicious behavior from transient heterogeneity, support post-hoc attribution, and extend forensic traceability to broader federated settings, attack families, adaptive adversaries, and pretrained or foundation-model adaptation.