Detection of false data injection and black-box adversarial attacks in peer-to-peer energy trading: A cloud-edge collaborative context-aware learning approach

Peer-to-peer (P2P) energy trading faces cybersecurity challenges, particularly false data injection (FDI) and black-box adversarial attacks. Existing monolithic security paradigms struggle to address these dual threats while meeting low-latency and high-concurrency requirements, and managing their asymmetric computational demands. This study proposes a novel Cloud-Edge Collaborative Learning Approach (CECLA) for adaptive defense. At the edge layer (L1), a meta-learning-tuned lightweight RealMLP detector achieves efficient FDI detection with sub-20-millisecond latency and over 95% accuracy, enhanced by SHAP for real-time decision explanation. At the cloud layer (L2), a Transformer-based TabICL model serves as an selective meta-detector, leveraging contextual learning to identify black-box adversarial attacks against L1 without retraining, achieving 93%–97% detection rates. L2 continuously adapts to emerging adversarial attacks through dynamic context sample updates. Experiments demonstrate the framework achieves high accuracy, low latency, and strong robustness on real-world energy data, providing a reliable security solution for next-generation P2P energy trading systems.