Research map / Adversarial robustness and security / Certified and empirical robustness
Improving VAEs' Robustness to Adversarial Attack
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TC regularization to a VAE | VAE · Chairs, 3D faces, and CelebA adversarial attacks | attack robustness | Helped |
| Use a hierarchical VAE with TC regularization and a likelihood conditioned on all latent layers | β-TCVAE · Chairs, 3D faces, and CelebA adversarial attacks | attack robustness | Helped |
| Add TC regularization | VAE downstream of attacked embeddings · SVHN, MLP on reconstructed image | accuracy: 0.17 -> 0.22 | Helped |
| Use hierarchical TC regularization | VAE downstream of attacked embeddings · SVHN, MLP on reconstructed image | accuracy: 0.22 -> 0.35 | Helped |
| Use hierarchical TC regularization | VAE downstream of attacked embeddings · SVHN, Conv classifier on reconstructed image | accuracy: 0.36 -> 0.41 | Helped |
| Use hierarchical TC regularization | VAE downstream of attacked embeddings · SVHN, MLP on latent representation | accuracy: 0.46 -> 0.57 | Helped |
| Use hierarchical TC regularization | VAE downstream of attacked embeddings · CIFAR10, MLP on reconstructed image | accuracy: 0.25 -> 0.38 | Helped |
| Use hierarchical TC regularization | VAE downstream of attacked embeddings · CIFAR10, Conv classifier on reconstructed image | accuracy: 0.32 -> 0.34 | Helped |
| Use hierarchical TC regularization | VAE downstream of attacked embeddings · CIFAR10, MLP on latent representation | accuracy: 0.26 -> 0.39 | Helped |
| Apply TC regularization to a chain-factorized DLGM whose data likelihood depends only on the bottom latent layer (ablation) | Hierarchical VAE · Adversarial attacks | attack robustness | No clear effect |
| Condition the data likelihood on all latent layers rather than only the bottom layer (ablation) | TC-regularized hierarchical VAE · Adversarial attacks | attack robustness | Helped |
| Increase β from the standard VAE setting (ablation) | β-TCVAE · dSprites, Chairs, and 3D faces latent-space attacks | attack loss | Mixed |
Closest papers by meaning
- Trading off Image Quality for Robustness is not Necessary with Regularized Deterministic Autoencoders (NeurIPS 2022 · similarity 0.93)
- Alleviating Adversarial Attacks on Variational Autoencoders with MCMC (NeurIPS 2022 · similarity 0.91)
- Adversarial Symmetric Variational Autoencoder (NeurIPS 2017 · similarity 0.90)
- Support is All You Need for Certified VAE Training (ICLR 2025 · similarity 0.89)
- Purify Unlearnable Examples via Rate-Constrained Variational Autoencoders (ICML 2024 · similarity 0.88)
- Why do Variational Autoencoders Really Promote Disentanglement? (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.