Adversarial Robustness of Deep Reinforcement Learning Detectors Against Stealthy Data Manipulation in Smart Meters

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

The integration of smart meters into modern energy distribution systems has improved grid efficiency but introduced critical vulnerabilities to false data injection (FDI) attacks that can manipulate consumption measurements for financial gain or grid destabilization. Deep reinforcement learning (DRL) detectors have emerged as adaptive solutions capable of identifying evolving attack patterns, yet their susceptibility to adversarial manipulation remains inadequately characterized. This study addresses this gap by developing a hybrid framework combining DRL-based detection, dynamic game-theoretic modeling, and blockchain verification to enhance robustness against stealthy data manipulation. Using a multi-head Graph Attention Network with LSTM for spatial-temporal feature extraction and a soft actor-critic DRL detector trained on real smart meter consumption data, the proposed system achieved 89.4% detection accuracy under adversarial conditions, with a true positive rate of 96.75% and false positive rate of 0.36%. The dynamic game model, formulated as a tri-level Stackelberg framework, reduced successful attack probability by 34% compared to static defense baselines. Blockchain verification via a two-tier reputation-based architecture provided immutable audit trails and malicious node detection, achieving the lowest attack success probability among benchmarked solutions. The findings demonstrate that integrating game-theoretic anticipation with blockchain verification substantially improves DRL detector resilience, offering a replicable framework for securing advanced metering infrastructure against sophisticated stealth attacks.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.