Inverted Generator Classifier Accurate and robust gradient-descent based classifier

In the field of deep learning, a traditional classifier receives input data and passes through hidden layers to output predicted labels. Conditional generators such as Conditional VAE [1] and Conditional GAN [2] receive latent vector and condition vector and generate data with the desired conditions. In this paper, I propose an Inverted Generator Classifier that uses conditional generators to find a pair of condition vectors and latent vectors that generate the data closest to the input data, and predict the label of the input data. Inverted Generator Classifier uses a trained conditional generator as it is. The inverted generator classifier repeatedly performs gradient descent by taking the latent vector for each condition as a variable and the model parameter as a constant to find the data closest to the input data. Then, among the data generated for each condition, the condition vector of the data closest to the input data becomes the predicted label. Inverted Generator Classifier is slow to predict because prediction is based on gradient descent, but has high accuracy and is very robust against adversarial attacks [3] such as noise. It is also not subject to gradient-descent based white-box attacks like FGSM [4]. Abbreviations Inverted Generator Classifier: IGC 1. Inverted Generator Classifier