Adversarial Patch Attacks Against Deep-Learning Based UAV Detection

The increased use of Unmanned Aerial Vehicles (UAVs) in both military and civilian domains has pushed the development of deep learning-based systems for automated detection. While effective these systems are vulnerable to adversarial attacks that can undermine their reliability. This thesis explores the design and implementation of adversarial patch attacks intended to evade state of the art detectors. The research begins by evaluating existing adversarial patches from the literature, finding their effec- tiveness to be limited when transferred to new models, which underscores the need for a custom tailored approach. This motivates the development of a white-box attack strategy targeting the YOLO family of object detectors (v5, v8, and v10). Initial experiments revealed a critical insight: an adversarial signal’s efficacy is dramatically enhanced when applied as a repeating, tiled pattern across the object’s surface, rather than as an isolated patch. This principle forms the basis of the primary contribution: an advanced generator that optimizes a base patch element with a masked, repeating pattern, effectively creating an adversarial camouflage. The re- sulting textures reduce the detection capabilities of all targeted YOLO models. Quantitative analysis shows these patterns achieve increasing Attack Success Rate for system configured with higher confidence detec- tion threshold. The findings demonstrate a practical and replicable method for creating and testing effective adversarial camouflage by analyzing this significant vulnerability in modern AI-based surveillance.