2019/12/06 by Sven Gowal, Chongli Qin, Gowal, Sven +10 · 1 citation
Computer Science · #Adversarial Robustness in Machine Learning #Generative Adversarial Networks and Image Synthesis #Anomaly Detection Techniques and Applications
paper · pdf · doi:10.48550/arxiv.1912.03192
Recent research has made the surprising finding that state-of-the-art deep\nlearning models sometimes fail to generalize to small variations of the input.\nAdversarial training has been shown to be an effective approach to overcome\nthis problem. However, its application has been limited to enforcing invariance\nto analytically defined transformations like \ℓp-norm bounded\nperturbations. Such perturbations do not necessarily cover plausible real-world\nvariations that preserve the semantics of the input (such as a change in\nlighting conditions). In this paper, we propose a novel approach to express and\nformalize robustness to these kinds of real-world transformations of the input.\nThe two key ideas underlying our formulation are (1) leveraging disentangled\nrepresentations of the input to define different factors of variations, and (2)\ngenerating new input images by adversarially composing the representations of\ndifferent images. We use a StyleGAN model to demonstrate the efficacy of this\nframework. Specifically, we leverage the disentangled latent representations\ncomputed by a StyleGAN model to generate perturbations of an image that are\nsimilar to real-world variations (like adding make-up, or changing the\nskin-tone of a person) and train models to be invariant to these perturbations.\nExtensive experiments show that our method improves generalization and reduces\nthe effect of spurious correlations (reducing the error rate of a "smile"\ndetector by 21% for example).\n