Secure Aggregation in Federated Learning for Non-IID Data Environments

Abstract

Training machine learning models across many institutions without moving raw data is attractive in theory, but the standard Federated Learning setup hands enormous power to one central aggregation server that participants must trust blindly. what happens when that trust assumption is violated and build a system, that does not require it. couples a four-layer distributed architecture with three stacked privacy mechanisms Local Differential Privacy to bound what any single gradient reveals, Secure Multi-Party Computation via the SecAgg protocol so the aggregation server never sees individual updates, and Tendermint-BFT blockchain consensus to create a tamper-evident audit log that no single party controls. Our main technical contribution is a formal proof that the combination of LDP and SecAgg satisfies end-to-end $(\varepsilon, \delta)$-differential privacy, including the SecAgg masking step cannot degrade the DP guarantee already earned by LDP noise injection. On CIFAR-10 with a ResNet-18 under three Dirichlet non-IID splits tested on real cluster hardware, reached 78.3% accuracy at $\varepsilon=0.83$ against six baselines including FedProx and SCAFFOLD-two standard non-IID benchmarks that prior blockchain FL work has consistently omitted.