An intelligent intrusion detection system for cyber-physical systems using GAN-LSTM networks

Cyber-Physical Systems (CPS) face increasing cybersecurity threats, demanding advanced intrusion detection methods. This research proposes a novel GAN-LSTM hybrid model to enhance anomaly detection in CPS by addressing key limitations of traditional approaches, including class imbalance and temporal...

Full description

Saved in:
Bibliographic Details
Main Authors: Md Shakil Siddique, Md. Ashikur Rahman Khan, Ishtiaq Ahammad, Nishu Nath, Joysri Rani Das, Fardowsi Rahman
Format: Article
Language:English
Published: Elsevier 2025-06-01
Series:Franklin Open
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2773186325000714
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Cyber-Physical Systems (CPS) face increasing cybersecurity threats, demanding advanced intrusion detection methods. This research proposes a novel GAN-LSTM hybrid model to enhance anomaly detection in CPS by addressing key limitations of traditional approaches, including class imbalance and temporal dependency learning. The primary objectives are: (i) developing an adversarial learning framework where the generator synthesizes realistic attack patterns while the discriminator improves detection robustness, (ii) introducing a hybrid anomaly scoring mechanism combining reconstruction and discrimination loss, and (iii) validating performance on real-world CPS datasets (SWaT and WADI). The model achieves 87 % accuracy (SWaT) and 91 % accuracy (WADI), with precision reaching 93 % (SWaT) and 97 % (WADI)—demonstrating strong capability to minimize false alarms. Notably, it attains 99 % recall on SWaT, ensuring near-complete attack detection, though WADI recall remains lower (75 %) due to complex attack patterns. The balanced F1-scores (91 % SWaT, 82 % WADI) outperform state-of-the-art methods like MAD-GAN by 14–45 %. Key innovations include LSTM-based temporal feature learning and GAN-driven synthetic minority oversampling, effectively handling CPS-specific challenges such as multivariate time-series complexity and rare attack instances. These results highlight the model’s potential for real-world CPS security, while future work will address deployment constraints like computational latency.
ISSN:2773-1863