Transformer-Based Intrusion Detection Systems: a Deep Federated Learning Approach for Privacypreserving Cybersecurity

Abstract

In the evolving landscape of cybersecurity, traditional intrusion detection systems (IDS) face significant challenges in handling high-dimensional data, real-time threat detection, and maintaining data privacy. To address these limitations, this paper proposes a novel TransformerBased Intrusion Detection System (IDS) integrated within a Deep Federated Learning (DFL) framework, aiming to achieve robust cybersecurity with strong privacy preservation. The proposed model leverages the selfattention mechanisms of transformers to effectively capture complex temporal and spatial dependencies inherent in network traffic, enabling highly accurate anomaly and attack detection. Meanwhile, federated learning ensures that sensitive data remains decentralized, minimizing privacy risks while collaboratively improving the global IDS model across distributed nodes. The system is trained and evaluated on multiple benchmark cybersecurity datasets, demonstrating superior performance compared to traditional convolutional and recurrent architectures. Experimental results reveal substantial improvements in detection accuracy, reduced false positive rates, and enhanced adaptability to emerging cyber threats. This work presents a scalable and privacy-preserving paradigm, opening new possibilities for next-generation IDS solutions in decentralized and sensitive environments such as healthcare, finance, and smart grids.