Increased-Confidence Adversarial Examples for Deep Learning Counter-Forensics

Transferability of adversarial examples is a key issue to apply this kind attacks against multimedia forensics (MMF) techniques based on Deep Learning (DL) in real-life setting. Adversarial example transferability, fact, would open the way deployment successful counter also cases where attacker does not have full knowledge to-be-attacked system. Some preliminary works shown that CNN-based image detectors are general non-transferrable, at least when basic versions implemented most popular libraries adopted. In paper, we introduce strategy increase strength and evaluate their transferability such varies. We experimentally show that, way, attack can be largely increased, expense larger distortion. Our research confirms security threats posed by existence even scenarios, thus calling for new defense strategies improve DL-based MMF techniques.