Adversarial Illusions in Machine Perception: An Imitation Game Approach
A new paper on arXiv proposes a disillusion paradigm for adversarial attacks in machine perception, framing the problem as an imitation game. The authors identify two types of adversarial illusions: deductive illusion, which crafts stimuli based on the victim model's decision logic, and inductive illusion, which embeds a backdoor during learning. The proposed unified defense framework addresses vulnerabilities across attack forms. The paper is available at arXiv:2501.19143.
Key facts
- Paper proposes a disillusion paradigm based on imitation game concept.
- Two types of adversarial illusions: deductive and inductive.
- Deductive illusion exploits decision boundaries to interfere with decision-making.
- Inductive illusion embeds a backdoor during learning phase.
- Unified defense framework addresses multiple attack forms.
- Paper available on arXiv with ID 2501.19143.
- Focus is on machine perception and adversarial attacks.
- Published as arXiv preprint.
Entities
Institutions
- arXiv