Study of Adversarial Machine Learning for Enhancing Human Action Classification

Action classification has traditionally been an effective way of identifying threatening and nonthreatening targets with significant traffic in military and civilian locations. Motivated actors in place of accessing detailed model requirements can simulate adversarial attacks on an action recognition networks that are highly transferable to production models. The problem is that adversarial attacks against spatial-temporal data are straightforward to generate since they are sequential and hard to detect. We propose using maximal informative adversarial training to create robust classification using Human action classification networks (HCNs). Using adversarial samples of the real action data, we can create synthetic sequences that expand the range of motions our classifier expects for a given class of actions. Our research demonstrates the effectiveness of adversarial trained HCNs against the transferred adversarial attacks. Our method of adversarial training using maximum informativeness outperformed defense by inference when comparing models that were transfer attacked using sequences of motions that are effective on previous models. We find that non-contextual human action recognition architectures are more susceptible than those using context from the scenes they are analyzing. Our contextual model performed 47 percent better after adversarial training. Contextual followed by graph-based models had the lowest attack success rate of all surrogate models trained.