GenAI-FDIA: Physics-Informed Generative Models for False Data Injection Attacks
A new framework, GenAI-FDIA, benchmarks 20 generative architectures for synthesizing physics-compliant false data injection attacks (FDIAs) on power systems. The models include Wasserstein GANs, MMD-VAEs, normalizing flows, diffusion models, and cross-family hybrids. Evaluated on IEEE 14-bus DC, 30-bus DC, and 14-bus AC testbeds with a 60/20/20 chronological split, all architectures achieved evasion rates of at least 86.6% against data-driven bad data detection (BDD) on the 14-bus network. The study also shows that limiting an attacker's topological knowledge reduces attack effectiveness. The work addresses data scarcity in training FDIA detectors by generating high-fidelity attacks that respect network physics.
Key facts
- GenAI-FDIA benchmarks 20 generative architectures for FDIA synthesis.
- Models include Wasserstein GANs, MMD-VAEs, normalizing flows, diffusion models, and hybrids.
- Evaluated on IEEE 14-bus DC, 30-bus DC, and 14-bus AC testbeds.
- 60/20/20 chronological split used for training, validation, and testing.
- All architectures achieved evasion rates ≥ 86.6% on the 14-bus network.
- Limiting topological knowledge reduces attack effectiveness.
- Addresses data scarcity in power system FDIA detection.
- Published on arXiv with ID 2605.18873.
Entities
Institutions
- arXiv
- IEEE