Inverting Adversarially Robust Networks for Image Synthesis

Renán A. Rojas-Gómez, Raymond A. Yeh, Minh N. Do, Anh Nguyen; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2221-2238


Despite unconditional feature inversion being the foundation of many image synthesis applications, training an inverter demands a high computational budget, large decoding capacity and imposing conditions such as autoregressive priors. To address these limitations, we propose the use of adversarially robust representations as a perceptual primitive for feature inversion. We train an adversarially robust encoder to extract disentangled and perceptually-aligned image representations, making them easily invertible. By training a simple generator with the mirror architecture of the encoder, we achieve superior reconstruction quality and generalization over standard models. Based on this, we propose an adversarially robust autoencoder and demonstrate its improved performance on style transfer, image denoising and anomaly detection tasks. Compared to recent ImageNet feature inversion methods, our model attains improved performance with significantly less complexity.

Related Material

[pdf] [supp] [code]
@InProceedings{Rojas-Gomez_2022_ACCV, author = {Rojas-G\'omez, Ren\'an A. and Yeh, Raymond A. and Do, Minh N. and Nguyen, Anh}, title = {Inverting Adversarially Robust Networks for Image Synthesis}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {2221-2238} }