Maciej Żelaszczyk

Maciej Żelaszczyk, Jacek Mańdziuk

In spite of the enormous success of neural networks, adversarial examples remain a relatively weakly understood feature of deep learning systems. There is a considerable effort in both building more powerful adversarial attacks and designing methods to counter the effects of adversarial examples. We propose a method to transform the adversarial input data through a mixture of generators in order to recover the correct class obfuscated by the adversarial attack. A canonical set of images is used to generate adversarial examples through potentially multiple attacks. Such transformed images are processed by a set of generators, which are trained adversarially as a whole to compete in inverting the initial transformations. To our knowledge, this is the first use of a mixture-based adversarially trained system as a defense mechanism. We show that it is possible to train such a system without supervision, simultaneously on multiple adversarial attacks. Our system is able to recover class information for previously-unseen examples with neither attack nor data labels on the MNIST dataset. The results demonstrate that this multi-attack approach is competitive with adversarial defenses tested in single-attack settings.

Maciej Żelaszczyk, Jacek Mańdziuk

Cross-modal representation learning allows to integrate information from different modalities into one representation. At the same time, research on generative models tends to focus on the visual domain with less emphasis on other domains, such as audio or text, potentially missing the benefits of shared representations. Studies successfully linking more than one modality in the generative setting are rare. In this context, we verify the possibility to train variational autoencoders (VAEs) to reconstruct image archetypes from audio data. Specifically, we consider VAEs in an adversarial training framework in order to ensure more variability in the generated data and find that there is a trade-off between the consistency and diversity of the generated images - this trade-off can be governed by scaling the reconstruction loss up or down, respectively. Our results further suggest that even in the case when the generated images are relatively inconsistent (diverse), features that are critical for proper image classification are preserved.