A Unified Lightweight Benchmark for Privacy-Preserving Federated Learning in Cyber-Physical Systems (Fashion-MNIST Case Study)

Federated learning (FL) enables collaborative training on decentralized data, but sharing model updates exposes clients to inference attacks. We introduce a lightweight, open benchmark for an apples-to-apples comparison of orthogonal privacy-enhancing technologies (PETs) under identical, resource-constrained conditions. While our framework supports homomorphic encryption (HE), this study focuses empirically on sample-level differential privacy (DP) and secure aggregation (SA) on a non-IID Fashion-MNIST setup with a shared CNN and optimizer.Across identical settings, we log model utility, communication, computation, and formal privacy. We observe a clear trade-off: the non-private FedAvg baseline reaches 86.50 % test accuracy in 6 rounds to convergence and 202.00 MB total communication. DP (final (ε, δ) = (1.86, 10−5)) drops to 76.30 % accuracy, requires 120 rounds to convergence, and transmits 4.05 GB. SA preserves high utility (82.50 %) with 7 rounds to convergence and 124.00 MB total traffic, adding only modest cryptographic overhead while shielding individual updates from a curious server. Taken together, our unified benchmark and results pro-vide actionable guidance for privacy choices in cyber-physical deployments where latency, bandwidth, and energy are tight.