Lightweight Differential Privacy for Federated Learning on Consumer AIoT Devices

Consumer Artificial Intelligence of Things (AIoT) endpoints, such as smart speakers, wearables, and home-automation hubs generate increasingly sensitive personal data, yet their severe constraints on memory, compute, and energy make rigorous on-device privacy a largely unsolved challenge. Federated Learning (FL) keeps raw data local, but gradient inversion and membership inference attacks reveal that vanilla FL provides no meaningful privacy guarantee. This paper proposes RDP-Fed, a Resource-Aware Differentially Private Federated Learning Framework engineered for the sub-1 GHz, kilobytes-of-RAM reality of consumer AIoT hardware. RDP-Fed tightly couples Ŕenyi Differential Privacy (RDP) accounting with adaptive Gaussian noise calibration and privacy-safe gradient sparsification, yielding a favorable privacy-utility-efficiency operating point that standard DP mechanisms cannot reach on constrained silicon. A heterogeneity-aware aggregation protocol further addresses the non-IID data distributions that characterize real consumer deployments. Evaluated on keyword spotting, human activity recognition, and predictive maintenance benchmarks, RDP-Fed reduces per-round communication overhead by up to 62% and energy consumption by 42% compared with DP-FedAvg, while holding model accuracy within 0.5 percentage points of the non-private federated baseline (FedAvg) and within 2.2 percentage points of the centralized training upper bound, all while satisfying formal differential privacy guarantees across all evaluated scenarios. These results position RDP-Fed as a practical, deployable privacy layer for the next generation of intelligent consumer electronics.