Topology-Aware Differential Privacy in Federated Learning
The episode discusses a paper titled "Topology-Aware Differential Privacy in Federated Learning." The authors introduce TADI, a framework that empirically breaks down data leakage into measurable channels based on structure and parameters. They also propose FULCRUM, a tailored noise allocation strategy that improves upon standard DP-SGD by addressing the structural asymmetry found in real-world AI systems.