A Privacy-Preserving and Efficient Split Learning Approach for IoT Network Security

Abstract

With the rising sophistication of cyberattacks, Network Intrusion Detection Systems (NIDS) are critical for organizational security. Traditional centralized deep learning approaches for NIDS have challenges with data privacy and the potential of sensitive network information leaks. This study presents a privacy-preserving NIDS architecture that employs Split Learning (SL) techniques, which vary from federated learning in that the neural network is divided into segments situated between the client (where the data source is located) and a central server. In this paradigm, raw network traffic data remains on the local client, but processed (or shattered) data from the split layer is transmitted to the server for further analysis, ensuring data secrecy and compliance with privacy rules such as GDPR. Evaluations of benchmark datasets like UNSW-NB15 and CIC-IDS2017 concentrate on detection accuracy, computing overhead, and communication efficiency. The results show that the Split Learning-based NIDS achieves detection performance comparable to centralized systems while significantly decreasing privacy threats and hardware demands on client devices. This paper proposes a scalable, secure collaborative intrusion detection method that can perform in resource-constrained and privacy sensitive environments. Keywords: Network Intrusion Detection (NIDS), Split Learning, Data Privacy, Cybersecurity, Deep Learning, Distributed Computing.