Alleviating adversarial attacks via convolutional autoencoder

In order to defend adversarial attacks in computer vision models, the conventional approach arises on actively incorporate such samples into the training datasets. Nonetheless, the manual production of adversarial samples is painful and labor intensive. Here we propose a novel generative model: Convolutional Autoencoder Model to add unsupervised adversarial training, i.e., the production of adversarial images from the encoded feature representation, on conventional supervised convolutional neural network training. To accomplish such objective, we first provide a novel statistical understanding of convolutional neural network to translate convolution and pooling computations equivalently as a hierarchy of encoders, and sampling tricks, respectively. Then, we derive our proposed Convolutional Autoencoder Model with the `adversarial decoders' to automate the generation of adversarial samples. We validated our proposed Convolutional Autoencoder Model on MNIST dataset, and achieved the clear-cut performance improvement over the normal Convolutional Neural Network.