Domain-Adaptive Split Federated Learning for Efficient Intrusion Detection in Industrial IoT Networks

Abstract

The rapid proliferation of 6G-enabled Industrial Internet of Things (IIoT) devices has driven unprecedented advancements in ultra-low-latency wireless communications, edge intelligence, and AI-native mobile computing, enabling largescale industrial automation and real-time situational awareness. Despite these benefits, securing IIoT systems against sophisticated cyber intrusions remains a critical challenge due to device mobility, stringent latency constraints, resource limitations, heterogeneity, and strict data privacy requirements inherent in 6G environments. While Federated Learning (FL) has emerged as a privacy-aware approach, it often struggles with computational load and intermittent connectivity inherent to mobile environments. In parallel, Split Learning (SL) efficiently reduces clientside computation but faces scalability limitations. To address these challenges, this paper proposes a novel Split Federated Learning (SFL) framework that blends the strengths of FL and SL. The proposed architecture supports lightweight clientside processing and robust server-side aggregation tailored to mobile IIoT contexts. Extensive experiments demonstrate the effectiveness of SFL in terms of classification accuracy, reduced communication overhead, adaptability to client heterogeneity, and improved resilience against mobility-induced intermittent connectivity, making it particularly suitable for mobile industrial intrusion detection applications.