CyberGuard: An Adaptive AI-Driven Backdoor Attack Detection
Iram Arshad, Saeed Hamood Alsamhi, Yuansong Qiao, Brian Lee, Yuhang Ye · IEEE Access · 2025
Smart manufacturing, a core of Industry 4.0 revolution, leverages Artificial Intelligence (AI) to boost efficiency and productivity. However, this data-driven interconnectivity exposes manufacturing to sophisticated cyber-attacks, one of the well-known attack is a backdoor that silently impairs AI systems. Traditional defence mechanisms are often rigid and not tailored to the unique requirements of Industry 4.0. To address this, we introduceCyberGuard, a novel Adaptive Anomaly Detection (AAD) framework designed specifically for detecting backdoor attacks in smart manufacturing environments. Unlike methods focusing on direct attack recognition, proposedCyberGuardmodels normal system behavior and identifies deviations as potential threats. It utilizes a combination of Convolutional Neural Network (CNN) and Transformer AutoEncoder (TAE) to detect anomalies within feature maps. Further, extensive testing on multiple datasets confirms CyberGuard’s capability to detect subtle yet crucial irregularities. We have conducted a comprehensive experiment against multiple backdoor attacks and compared and evaluatedCyberGuardwith Neural Cleanse, STRIP, Active Clustering, and Spectral Signature. In addition, an additional experiment is also conducted by employing a T-test statistical analysis to demonstrate the significant differentiation between normal and compromised feature maps, thereby validatingCyberGuardeffectiveness without prior distribution knowledge of the attacks. The results demonstrate thatCyberGuardsignificantly outperformed traditional defences by achieving approximate 100% detection with high F1-score compared to existing defences.