Reiner N. Dizon-Paradis

Domain-specific knowledge, like standard cell library layouts, turns federated learning into a privacy nightmare for integrated circuit training data, enabling high-fidelity reconstruction attacks without auxiliary datasets.

FL's promise of privacy in hardware assurance crumbles as a novel data-free attack reconstructs training data from model updates, revealing sensitive hardware IP without needing auxiliary datasets.

Federated learning, while improving deep learning for hardware assurance, fails to protect sensitive intellectual property due to its vulnerability to gradient inversion attacks that can recover SEM images.