Transcoders: A Better Alternative To Denoising Autoencoders

Image denoising is a popular technique that is used to remove noise incurred due to hardware faults or noise carefully crafted by an attacker. Autoencoders are some of the most popular denoisers. Their ability to learn a distribution's latent space helps them achieve this property, and they are generally good at it. However, they are known to fumble in a white-box threat model where an attacker knows everything about the victim classifier and its denoiser network - including its architecture and hyperparameters. We show that this problem stems from the autoencoder's learning goal. In this paper, we augment an autoencoder's learning goal to conceive what we call transcoders. This modification forces the transcoder network to learn a function that is more adept at denoising a given image. Our results, evaluated on two datasets - MNIST and CIFAR10, a slew of attacks, and two threat models - gray-box and white-box, help us argue the following: given a denoising tool built using an autoencoder, one can update the learning goal of the autoencoder to that of a transcoder, and achieve a transcoder-based denoiser that is significantly better at handling both fault-induced and attack-induced noise.