Blockchain-Augmented FL IDS for Non-IID Edge-IoT Data Using Adaptive Trimmed Mean Aggregation

Abstract

The rapid expansion of the Internet of Things (IoT) into edge networks, which are populated by resource-constrained devices, introduces significant security challenges. This article introduces a robust intrusion detection systems (IDS) for edge-IoT networks developed as a combination of blockchain technology and federated learning (FL) with a customized aggregation strategy, adaptive trimmed mean aggregation (ATMA). Our design leverages a permissioned blockchain to authenticate clients and immutably store the final global model, guaranteeing that only verified participants contribute to the training. ATMA strategy used in the FL departs from fixed-threshold schemes in other works by dynamically adjusting its trimming parameter according to the observed variance in client updates. This variance-aware trimming provides strong Byzantine resilience without sacrificing model accuracy, and its sorting-based implementation maintains an O(n log n) computational complexity. The proposed setup was evaluated under combined label-flipping and Gaussian-noise attacks at adversarial rates of 0%, 10%, 20%, 30%, 40%, 50%, and 60%, in both IID and non-IID data distributions. The results demonstrated that our blockchain-backed ATMA preserves high detection performance under severe attack scenarios and does so with minimal overhead, making it a scalable, secure solution for safeguarding Edge-IoT deployments.