Backdoor Attacks Threaten Fault Detection in Cyber-Physical Systems
A new arXiv preprint (2605.27674) reveals that machine learning-based fault detection and localization frameworks in Cyber-Physical Systems (CPS) are vulnerable to backdoor attacks. CPS integrate sensing, communication, computation, and control for critical infrastructure like smart grids and industrial automation. In electrical utilities, controllers rely on ML/DL models to detect faults (e.g., voltage fluctuations) and perform load balancing. However, adversaries can inject malicious patterns into training data, causing models to behave normally until triggered by specific patterns, then output attacker-controlled results. The paper highlights this security gap without proposing a defense.
Key facts
- arXiv preprint 2605.27674 addresses backdoor attacks on CPS fault detection.
- CPS integrate sensing, communication, computation, and control.
- Critical infrastructure includes smart grids and industrial automation.
- Controllers in electrical utilities detect faults like voltage fluctuations.
- ML/DL frameworks are used for real-time anomaly detection.
- Backdoor attacks inject malicious patterns into training data.
- Attacked models behave normally until triggered by specific patterns.
- No defense is proposed in the paper.
Entities
Institutions
- arXiv