Cyber Sentry: Strengthening Security Infrastructure for Industrial Cyber-Physical Systems Using Federated Deep Learning

Abstract

Industrial Cyber-Physical Systems (ICPS) represent the backbone of applications, such as manufacturing, energy, and healthcare, but they are also under the threat of advanced cyberattacks, e.g., zero-day and data leak ones. The shortcomings of centralized IDS on the aspects of data security, privacy preserving and efficiency have been discovered. To address this challenge, we propose Cyber Sentry, a federated deep learning (DL) framework which enhances the security of ICPS by allowing decentralized, collaborative model training with no access to sensitive data. Data-centeredness: Here, the RT-IoT2022 dataset is vertically sliced, and dynamic pre-processing is achieved to train deep NNs, such as CNNs and LSTMs, in a local training fashion at edge devices. Based on those models, a strong global model for an anomalous situation is aggregated for detection. The framework is validated experimentally by achieving 92.5% detection accuracy with negligible false positive, while preserving the privacy of data through encryption mechanism. It has also been analyzed for enhancing the security at the edge layer by leaning on edge computing and blockchain security systems to achieve improved scalability and defense capabilities against cyber-attacks. It also shows advantages in terms of reduced communication cost and increased operation availability. We offer future research directions from an academic perspective and some implications for the industry on the adoption of federated learning to cybersecurity for ICPS. Some future work is to enhance the adversarial attack resistance, to integrate federated learning in blockchain networks, and to explore how to implement explainable AI to make the model more explainable. The quality of the experimental result offered by the proposed method demonstrates the need for federated deep learning to protect the industrial infrastructures in a connected world.