Initialization of Random Vectors to Enhance Defense-GAN for Image Classification

Deep learning has been successfully applied to a wide spectrum of applications. Nevertheless, its performance is seriously degraded by adversarial attacks, an elaborately designed small perturbation to input data that can mislead the model prediction. Specifically, life-critical deployments of deep learning such as autonomous vehicles are a major concern of these attacks. Defense-GAN is considered a state-of-the-art defense method that is robust to both black-box and white-box adversarial attacks. Its defense framework generates new data close to the unperturbed ones from a given adversarial sample before feeding to the predictive model. But Defense-GAN takes time to determine reconstructed data at inference time due to the iterative nature of gradient descent. In this work, the improvement of Defense-GAN to rapidly generate well-reconstructed data from adversarial samples is proposed using carefully crafted initialization of random vectors. The experimental results demonstrate that the proposed method can outperform both classification performance and inference time compared with the Defense-GAN on the benchmark datasets.