Correction-based Defense Against Adversarial Video Attacks via Discretization-Enhanced Video Compressive Sensing

We introduce SECVID, a correction-based framework that defends video recognition systems against adversarial attacks without prior adversarial knowledge. It uses discretization-enhanced video compressive sensing in a black-box preprocessing module, transforming videos into a sparse domain to disperse and neutralize perturbations. While SECVID's discretized compression disrupts perturbation continuity, its reconstruction process minimizes adversarial elements, causing only minor distortions to the original videos. Though not completely restoring adversarial videos, SECVID significantly enhances their quality, enabling accurate classification by SECVID-enhanced video classifiers and preventing adversarial attacks. Tested on C3D and I3D with the UCF-101 and HMDB-51 datasets against five types of advanced video attacks, SECVID outperforms existing defenses, improving detection accuracy by 38.5% to 866.2%. Specifically designed for high-risk environments, SECVID addresses trade-offs like minor accuracy reduction, additional pre-processing training, and longer inference times, with potential optimization through selective security impacting strategies.