KAUST DepartmentKing Abdullah University of Science and Tehchnology (KAUST), Thuwal, Saudi Arabia
Electrical and Computer Engineering Program
Computer, Electrical and Mathematical Science and Engineering (CEMSE) Division
Computer Science Program
Visual Computing Center (VCC)
KAUST Grant NumberOSR-CRG2019-4033
Embargo End Date2023-10-24
Permanent link to this recordhttp://hdl.handle.net/10754/677988
MetadataShow full item record
AbstractThis work evaluates the robustness of quality measures of generative models such as Inception Score (IS) and Fréchet Inception Distance (FID). Analogous to the vulnerability of deep models against a variety of adversarial attacks, we show that such metrics can also be manipulated by additive pixel perturbations. Our experiments indicate that one can generate a distribution of images with very high scores but low perceptual quality. Conversely, one can optimize for small imperceptible perturbations that, when added to real world images, deteriorate their scores. We further extend our evaluation to generative models themselves, including the state of the art network StyleGANv2. We show the vulnerability of both the generative model and the FID against additive perturbations in the latent space. Finally, we show that the FID can be robustified by simply replacing the standard Inception with a robust Inception. We validate the effectiveness of the robustified metric through extensive experiments, showing it is more robust against manipulation.
CitationAlfarra, M., Pérez, J. C., Frühstück, A., Torr, P. H. S., Wonka, P., & Ghanem, B. (2022). On the Robustness of Quality Measures for GANs. Computer Vision – ECCV 2022, 18–33. https://doi.org/10.1007/978-3-031-19790-1_2
SponsorsThis work was supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research (OSR) under Award No. OSR-CRG2019-4033.
PublisherSpringer Nature Switzerland
Conference/Event name17th European Conference on Computer Vision, ECCV 2022