Adversarial Images Through Steganographic Glasses

A study explored the connection between steganography and adversarial images, finding that steganography helps attackers more than defenders. Steganalysis helps detect adversarial perturbations, while steganography can forge adversarial perturbations that are invisible to the human eye and statistically undetectable. The study applied information-hiding tools to attacking or defending computer vision image classification, and conducted attack-defense experiments on advanced image classifiers, steganalyzers, and steganographic embedding schemes.