Adversarially-trained encryption models to resist AI-based image reconstruction attacks

The recent years have seen the fast development of deep learning and the associated big leap in image reconstruction quality, allowing adversaries to extract high-value visual information (such as photos of people) form encrypted or obfuscated data. Conventional encryption methods are not always sufficient to withstand advanced AI-powered attacks, especially the ones that involve the use of generative adversarial networks (GANs) or inversion models. In this paper we propose a new adversarially-trained encryption scheme that is resistant against such image reconstruction attacks. The method is based on a min-max game between an encryption network and a simulated attacker network, such that encrypted images have visual irreversibility but are computationally light-weight. Thorough experiments on standard image benchmarks, such as CIFAR-10 and Celeb A, show that our technique can greatly decrease the success probability of AI-based reconstruction without hurting the performance of encryption. The findings point to the promise of adversarial training as an effective technique to ensure the visual data security in the age of intelligent threats.