Explicit Optimization of min max Steganographic Game

This article proposes an algorithm which allows Alice to simulate the game played between her and Eve. Under condition that set of detectors assumes Eve have is sufficiently rich (e.g. CNNs), she has enabling avoid detection by a single classifier (e.g adversarial embedding, gibbs sampler, dynamic STCs), proposed converges efficient steganographic algorithm. possible using min max strategy consists at each iteration in selecting least detectable stego image for best among Eve's learned classifiers. The extensively evaluated compared prior arts results show potential increase practical security classical methods. For example error probability Perr XU-Net on detecting images with payload 0.4 bpnzAC embedded J-Uniward QF 75 starts 7.1% increased +13.6% reach 20.7% after eight iterations. same embedding rate 95, undetectability 23.4%, it jumps +25.8% 49.2% 3.