The School of Computer Science is pleased to present…
Forensic-Ready Federated Learning: Evidence-Producing Frameworks for Accountable Training and Verifiable Unlearning
PhD Dissertation Proposal by: Safiia Mohammed
Date: November 27, 2026
Time: 10 am to 12 pm
Location: Room 122 Essex Hall
Abstract:
Federated learning (FL) trains shared models without pooling raw data, but its decentralization weakens accountability: after poisoning, backdoor insertion, or an unlearning request, investigators often cannot reconstruct events, verify evidence integrity, or attribute actions to a responsible party.
This dissertation proposes designing FL as an evidence-producing system across its lifecycle. An accountability-oriented survey of poisoning attacks in FL identifies gaps in forensic attribution, malicious-server threat models, and evidence governance. FL-Forensics, a forensic-ready extension of Flower, combines signed update commitments, independent witnessing, hash-chained ledgers, and blockchain anchoring to produce tamper-evident audit trails, evaluated with a new Forensic Quality Index (FQI). Forensic Federated Unlearning (FFU) separates model correction from forensic assurance and certifies unlearning only when execution and outcome checks pass, reducing post-unlearning attack success by 85–95% over a verification-oriented baseline. Remaining work will unify these components into an end-to-end forensic lifecycle and evaluate it under secure aggregation, larger federations, and adaptive adversaries.
Keywords: Federated Learning; AI Forensics; Federated Unlearning; Accountability.
Thesis Committee:
Internal Reader: Dr. Imran Ahmad
Internal Reader: Dr. Luis Rueda
External Reader: Dr. Abdulkadir Hussein
Advisor: Dr. Alioune Ngom
