FlowGuard Stops AI Model Theft Attacks on Power Grid Defenses
FlowGuard uses generative flow models to catch model-stealing attacks on energy grid AI, surviving distributed Sybil attacks where existing defenses collapse to 0% detection
Takeaways
- Model theft threat: AI intrusion detection systems in energy grids can be cloned via data-free query attacks (MAZE, DisGUIDE), enabling offline evasion.
- FlowGuard defense: Uses a Continuous Normalizing Flow to score each query's log-likelihood; synthetic attack queries score low and are blocked before reaching the IDS.
- Key insight: Synthetic queries occupy a lower-dimensional manifold than real traffic, making them detectable via density estimation without identity tracking.
- Sybil-proof: PRADA's detection rate dropped to 0% against 100-client distributed attacks; FlowGuard maintained stable detection in both single and distributed settings.
- Hard-label compatible: Works where prediction-poisoning defenses cannot, since it operates on inputs rather than outputs.
- Limitation: Currently only effective against data-free attacks; data-dependent attacks where the adversary holds real traffic remain an open problem.
AI-powered intrusion detection systems (IDS) are now a frontline defense for power grids, smart meters, and energy infrastructure. But there is a quiet threat that most defenders have not accounted for: an attacker who does not try to break in directly, but instead steals the IDS model itself. With a cloned model in hand, they can craft traffic that evades detection entirely, offline, with no risk of being caught in the act. A new paper proposes FlowGuard, a defense that detects these theft attempts at the query level, without needing to know who is asking.
The threat nobody is defending against
AI-based Intrusion Detection Systems deployed in energy infrastructure are vulnerable to model theft attacks, which allow adversaries to create evasive traffic offline. The attack class in question is called model extraction or model stealing: an adversary queries the IDS repeatedly, collects its predictions, and trains a clone model that mimics the original. Once they have the clone, they can probe it locally to find inputs that slip past the real detector.
What makes this especially dangerous in the energy sector is the data-free variant of the attack. Recent model-extraction attacks on Machine Learning as a Service systems have moved towards data-free approaches, showing the feasibility of stealing models trained with difficult-to-access data. Attacks like MAZE and DisGUIDE do not need any real network traffic to start. MAZE guides a generator toward regions of maximal model disagreement via zeroth-order gradient estimation, while DisGUIDE introduces a novel generator training scheme that maximizes the disagreement loss between two clone models that attempt to copy the model under attack. The result is a fully synthetic query stream that can extract a working clone without ever touching real data.
Why existing defenses fall apart
Current defenses against model extraction rely either on identity-bound query monitoring, which is ineffective against distributed attackers (Sybil), or on prediction poisoning through soft-label perturbation, which is inapplicable to hard-label IDS deployments. To understand why this matters, consider each failure mode separately.
- PRADA is the canonical detection-based defense. PRADA detects malicious queries by assessing the statistical distribution of the minimum L2-norm distance between queries. It works by tracking query history per user. But that means it relies on identity: if an attacker splits their queries across 100 fake clients (a Sybil attack), each individual client looks perfectly innocent. PRADA's detection rate dropped to 0% when the distribution changed.