Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Adversarial Distillation of Bayesian Neural Network Posteriors
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 |
|---|---|---|---|
| Distill SGLD posterior samples with APD | BNN classification on MNIST · MNIST, fcNN1 and fcNN2 | accuracy: 0.979 -> 0.978; 0.980 -> 0.981 | No clear effect |
| Use APD instead of MC-Drop | MNIST anomaly detection · notMNIST, OmniGlot, CIFAR10bw, Gaussian, Uniform; fcNN2, VR | AUROC: 88.0 -> 97.8; 91.5 -> 98.8; 90.1 -> 96.9; 91.3 -> 99.6; 93.6 -> 99.8 | Helped |
| Distill SGLD posterior samples with APD | MNIST anomaly detection · notMNIST and CIFAR10bw, fcNN2, VR | AUROC: 98.1 -> 97.8; 97.4 -> 96.9 | No clear effect |
| Use APD instead of MC-Drop | MNIST adversarial detection · FGSM and PGD transferred from SGLD | AUROC: 54.99 -> 75.93; 56.91 -> 82.80 | Helped |
| Distill SGLD posterior samples with APD | MNIST adversarial detection · FGSM and PGD transferred from SGLD | AUROC: 83.76 -> 75.93; 84.98 -> 82.80 | Hurt |
| Increase the number of test-time samples from the GAN (ablation) | APD · notMNIST anomaly detection | VR detection performance | Helped |
| Increase the number of Gaussian-mixture components (ablation) | MoG approximation of SGLD posterior · notMNIST anomaly detection | anomaly detection performance | Helped |
| Use 60-component diagonal-covariance MoG instead of SGLD samples (ablation) | Posterior sampler · notMNIST anomaly detection | performance relative to SGLD: 99.3% | No clear effect |
| Use WGAN-GP instead of the original GAN or weight-clipped WGAN (ablation) | APD · GAN training | convergence stability | Helped |
| Use 20 GAN-generated test-time samples instead of 20 SGLD samples (ablation) | APD · notMNIST anomaly detection | performance | Hurt |
| Generate more than 20 test-time samples from the GAN (ablation) | APD · notMNIST anomaly detection | performance: 50 SGLD samples; ~2.5x storage savings | Helped |
Closest papers by meaning
- Adv-BNN: Improved Adversarial Defense through Robust Bayesian Neural Network (ICLR 2019 · similarity 0.90)
- Improving Bayesian Neural Networks by Adversarial Sampling (AAAI 2022 · similarity 0.89)
- Bayesian GAN (NeurIPS 2017 · similarity 0.89)
- Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo (ICLR 2025 · similarity 0.88)
- A Study of Bayesian Neural Network Surrogates for Bayesian Optimization (ICLR 2024 · similarity 0.86)
- Robust Distillation via Untargeted and Targeted Intermediate Adversarial Samples (CVPR 2024 · similarity 0.85)
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.