Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 21·6 min read
GenAI-FDIA: Physics-Informed Generative Models for False Data Injection Attacks
TL;DR
Power grid operators lack realistic attack data to test intrusion detectors because real operational measurements are proprietary and manual attacks don't reflect actual network physics constraints. GenAI-FDIA benchmarks 20 generative model architectures—including Wasserstein GANs and MMD-VAEs (variational autoencoders using maximum mean discrepancy loss)—to synthetically generate physics-compliant false data injection attacks (FDIA: forged sensor readings that bypass grid safeguards).
✦ Why It Matters
Grid operators can now stress-test FDIA detectors with realistic synthetic attacks without exposing sensitive operational data or manually crafting implausible scenarios.
Key Takeaways
- Use GenAI-FDIA to generate realistic synthetic FDIA datasets respecting network physics constraints.
- Benchmark 20 generative architectures to identify best performers for your grid topology.
- Train detectors on synthetic data to avoid exposing proprietary operational measurements.
- Evaluate detector robustness against physics-compliant attacks, not hand-crafted unrealistic scenarios.
Full Summary
Power systems face false data injection attacks (FDIA)—malicious manipulation of sensor readings to disrupt grid operations. Training detectors for these attacks is bottlenecked by data scarcity: real operational measurements are commercially sensitive, and hand-crafted synthetic attacks fail to capture the complex statistical patterns imposed by underlying network physics.
GenAI-FDIA addresses this by systematically benchmarking 20 generative model architectures spanning Wasserstein GANs (adversarial networks optimized for distribution matching), MMD-VAEs (variational autoencoders using maximum mean discrepancy), and other approaches. These models learn to generate synthetic FDIA scenarios that respect physical constraints of power grids while remaining realistic.
The framework enables security researchers to train and evaluate attack detectors at scale without accessing proprietary grid data, significantly reducing the data scarcity bottleneck in power system security research.