Attacking Learning-based Models in Smart Grids: Explainability as a Double-Edged Sword

Smart grids increasingly rely on learning-based components for monitoring, control, and security-critical decision making. To address concerns regarding transparency, accountability, and regulatory compliance, eXplainable Artificial Intelligence (XAI) techniques are often integrated into these systems. However, the security implications of explainability in adversarial environments remain insufficiently understood. This dissertation investigates explainability from an attacker-centric perspective and analyzes how XAI alters the threat landscape of learning-based smart grid systems.

Adopting a proactive adversarial methodology, the thesis combines threat analysis, a review of existing attacks, and extensive empirical evaluations conducted in realistic smart grid testbeds. Across multiple use cases (including intrusion detection, power quality recognition and industrial vision pipelines) the work demonstrates that XAI can be systematically exploited to amplify adversarial attacks against integrity, availability, and confidentiality of targeted models. Explanation methods are shown to reduce attacker uncertainty, guide feature-space and problem-space perturbations, enable efficient data poisoning, and facilitate model extraction and covert data exfiltration.
... mehr

Beyond demonstrating attacks, the thesis explores defense directions that explicitly account for XAI-induced attack surfaces. A lightweight Moving Target Defense strategy is evaluated, showing that redistributing feature importance and eliminating brittle correlations can significantly improve robustness without incurring substantial overhead.

Overall, this work establishes explainability as a double-edged sword in smart grid security: while enhancing transparency for legitimate stakeholders, it simultaneously expands the attacker’s capability space. The findings highlight the need for security-aware explainability and adversarially informed design principles in future learning-based critical infrastructure systems.