Stealthy Backdoors as Compression Artifacts

Model compression is a widely-used approach for reducing the size of deep learning models without much accuracy loss, enabling resource-hungry to be compressed use on resource-constrained devices. In this paper, we study risk that model could provide an opportunity adversaries inject stealthy backdoors. backdoor attack machine model, adversary produces performs well normal inputs but outputs targeted misclassifications containing small trigger pattern. We design attacks such full-sized released by appears free from backdoors (even when tested using state-of-the-art techniques), it exhibits highly effective backdoor. show can done two common techniques—model pruning and quantization—even in settings where has limited knowledge how particular will done. Our findings demonstrate importance performing security tests actually deployed not their precompressed version. implementation available at https://github.com/yulongtzzz/Stealthy-Backdoors-as-Compression-Artifacts .