Work

Focus Matters: Advancing Adversarial Patch Transferability through Contextual Perception and Gradient-Prior Masking

Year: 2026
Type: article
Abstract: Adversarial patches pose a potent physical-world threat, relying on transferability from surrogate models to attack black-box detectors. Yet, this transferability is often crippled by two fundamental yet overlooked issues. First, object detectors exhibit a pronounced bias towards background features, which prevents adversarial patches from fully leveraging high-value foreground information, thereby constraining their optimization space and transferability. Second, redundant pixels within the patch dilute the attack by interfering with the optimization of pivotal ones. To address these challenges, we propose a Context-Enhanced Perception (CEP) method and a Prior-Guided Masking (PGM) strategy to significantly boost attack transferability. The CEP enhances the patch’s ability to perceive and exploit salient foreground features, compelling the model to shift its focus from distracting background information to the object of interest. This ensures that the patch can be optimized based on the most valuable features. The PGM implements a dynamic masking strategy that leverages gradient priors to identify and deactivate non-essential pixels. By focusing solely on key pixels to optimize the perturbation, it further maximizes patch aggressiveness. Extensive experiments demonstrate the superiority of our approach. In white-box and black-box settings, it yields larger decreases in average precision than the state of the art by 7.27% and 11.07%, respectively.
Language: English
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204138642
FWCI:
Cites:
Cited by:
Related to: 10