Adversarial Attack for Deep Steganography Based on Surrogate Training and Knowledge Diffusion

Deep steganography (DS), using neural networks to hide one image in another, has performed well terms of invisibility, embedding capacity, etc. Current steganalysis methods for DS can only detect or remove secret images hidden natural and cannot analyze modify content. Our technique is the first approach not effectively prevent covert communications DS, but also content communications. We proposed a novel adversarial attack method considering both white-box black-box scenarios. For attack, several loss functions were applied construct gradient- optimizer-based that could delete images. As more realistic case, was based on surrogate training knowledge distillation technique. All tested Tiny ImageNet MS COCO datasets. The experimental results showed completely even container while maintaining latter’s high quality. More importantly, be regarded as new approach.