Robustness of Machine Learning Based Compression

We compare in this work the robustness of machine learning based image compression algorithms with classical algorithms such as JPEG. For this, we run adversarial attacks against [2] and [1] as two examples for the first type of image compression networks, and a differentiable variant of JPEG [3] which is used for training the GAN attack, and measure the output error with a JPEG implementation. For training the attacking network, we follow an approach that is related, but not identical to that of [4]. The objective function is here