The Vanishing-Truth Patch: A Scale-Adaptive Vanishing Attack for Anchor-Based YOLO Detectors

Deep learning-based object detection is essential for aerial surveillance systems, yet it remains critically vulnerable to adversarial attacks. While adversarial patches can conceal objects, current attack methods often lack consistent effectiveness across varying object scales, which poses a fundamental limitation for their robustness in practical aerial scenarios. To address this challenge, we propose the Vanishing-Truth Attack, a novel adversarial patch framework that specifically optimizes a single patch to remain effective across continuous scale changes. Our approach integrates a geometry-aware adaptive scaling (GAAS) mechanism with a tailored objective function, enabling robust dynamic adaptation without requiring multiple patch instances. We evaluate our method on three anchor-based YOLO detectors: YOLOv3, YOLOv5s, and YOLOv5l. Experiments on the challenging multiscale MAR20 dataset demonstrate that our attack achieves a high success rate in concealing objects across all three models, outperforming the compared baseline methods. This work highlights a critical vulnerability in AI perception and provides foundational insights for developing more robust and certifiably secure AI systems.