Adaptive Lightweight Federated Learning With Aggregation-Only CKKS for Privacy-Preserving IoT Intrusion Detection

Federated learning (FL) enables collaborative training without sharing raw data, but standard FL exposes client updates and burdens resource-constrained IoT devices. The authors propose AdaptiveCKKS, an FL framework combining aggregation-only CKKS encryption with index-free block sparsification and stochastic quantization. A lightweight controller adaptively selects compression ratio and quantization per device/round based on on-device calibration of bandwidth, CPU, and encryption cost, while CKKS contexts are fixed at enrollment. The server performs ciphertext-only additions, decrypting only the aggregate each round. On BoT-IoT and ToN-IoT datasets, AdaptiveCKKS improves accuracy by 3.2–3.8% over FL and fixed-HE, reduces per-round communication by ~45% and average power by ~39%, and increases resistance to membership inference and gradient inversion attacks. Results are averaged over 10 runs with 95% confidence intervals, and all artifacts are released for reproducibility.