A study of the effect of JPG compression on adversarial images
This addresses a security and machine learning challenge for neural network image classifiers, but it is incremental as it builds on known adversarial vulnerabilities.
The study investigated how JPG compression affects adversarial images, finding that for small-magnitude Fast-Gradient-Sign perturbations, compression often largely reverses the drop in classification accuracy, but not always, and it becomes insufficient as perturbation magnitude increases.
Neural network image classifiers are known to be vulnerable to adversarial images, i.e., natural images which have been modified by an adversarial perturbation specifically designed to be imperceptible to humans yet fool the classifier. Not only can adversarial images be generated easily, but these images will often be adversarial for networks trained on disjoint subsets of data or with different architectures. Adversarial images represent a potential security risk as well as a serious machine learning challenge---it is clear that vulnerable neural networks perceive images very differently from humans. Noting that virtually every image classification data set is composed of JPG images, we evaluate the effect of JPG compression on the classification of adversarial images. For Fast-Gradient-Sign perturbations of small magnitude, we found that JPG compression often reverses the drop in classification accuracy to a large extent, but not always. As the magnitude of the perturbations increases, JPG recompression alone is insufficient to reverse the effect.