English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

In an unsupervised attack on variational autoencoders (VAEs), an adversaryfinds a small perturbation in an input sample that significantly changes itslatent space encoding, thereby compromising the reconstruction for a fixeddecoder. A known reason for such vulnerability is the distortions in the latentspace resulting from a mismatch between approximated latent posterior and aprior distribution. Consequently, a slight change in an input sample can moveits encoding to a low/zero density region in the latent space resulting in anunconstrained generation. This paper demonstrates that an optimal way for anadversary to attack VAEs is to exploit a directional bias of a stochasticpullback metric tensor induced by the encoder and decoder networks. Thepullback metric tensor of an encoder measures the change in infinitesimallatent volume from an input to a latent space. Thus, it can be viewed as a lensto analyse the effect of input perturbations leading to latent spacedistortions. We propose robustness evaluation scores using the eigenspectrum ofa pullback metric tensor. Moreover, we empirically show that the scorescorrelate with the robustness parameter $\beta$ of the $\beta-$VAE. Sinceincreasing $\beta$ also degrades reconstruction quality, we demonstrate asimple alternative using \textit{mixup} training to fill the empty regions inthe latent space, thus improving robustness with improved reconstruction.

Adversarial robustness of VAEs through the lens of local geometry