Sensitivity‑Aware Hardening and Run‑Time Detection of Stealthy Weight‑Drift Trojans in ML Accelerators

Title	Sensitivity‑Aware Hardening and Run‑Time Detection of Stealthy Weight‑Drift Trojans in ML Accelerators
Summary	design and evaluate sensitivity-aware defences that detect and mitigate stealthy, gradual weight-drift Trojans in FPGA-based ML accelerators
Keywords
TimeFrame
References	Grimsholm, Filip, and Cassandra Westergren. "Resilience of Machine Learning Hardware Accelerators Against Accuracy Degrading Trojans." (2025).
Prerequisites
Author
Supervisor	Mahdi Fazeli
Level	Master
Status	Open

In this thesis, you will design and evaluate sensitivity-aware defences that detect and mitigate stealthy, gradual weight-drift Trojans in FPGA-based ML accelerators. You’ll combine PD-guided “canary” monitors with selective integrity checks on the most critical weights, then measure detection speed, false alarms, and hardware cost. The work includes reproducing a baseline attack setup and delivering a practical, low-overhead defence recipe with code.