Concept-based Adversarial Attack: a Probabilistic Perspective

Researchers propose a new concept-based adversarial attack framework that targets entire concept distributions rather than single images, generating diverse adversarial examples while preserving the original concept identity. The method creates adversarial images with variations in pose, viewpoint, or background that can still mislead classifiers while remaining recognizable as instances of the original category.

• →New adversarial attack framework operates on concept distributions rather than individual images to generate diverse examples.
• →Method preserves original concept identity while creating variations in pose, viewpoint, and background.
• →Approach maintains mathematical consistency with traditional adversarial attack frameworks.
• →Concept-based attacks demonstrate higher attack efficiency compared to single-image perturbations.
• →Framework generates more diverse adversarial examples while effectively preserving underlying concepts.