On the Impact of Model Tolerance in Power Grid Anomaly Detection Systems

Abstract

Rapid development in deep learning-based detection systems for numerous industrial applications has opened opportunities to apply them in power grids. A consumer’s power consumption can be monitored to recognize any anomalous behavior in their household. When building such detection systems, evaluating their robustness to adversarial samples is critical. It has been shown that when we provide adversarial samples to deep learning models, they falsely classify instances, even when the perturbation or noise added to the original data is very small. On the other hand, these models should be able to detect attack instances correctly and raise few to no false alarms. While this expectation can be difficult to attain, we are allowed to choose a threshold that decides the extent to which the detection and false alarm rates are compromised. To this end, we explore the threshold selection problem for state-of-the-art deep learning-based detection models such that it can recognize attack instances. We show that selecting a threshold is challenging, and even if an appropriate threshold is chosen, the tolerance of a model to adversarial samples can still leave avenues for an attack to be successful.