Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Adversarial robustness of amortized Bayesian inference
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 |
|---|---|---|---|
| Apply targeted PGD perturbations rather than random perturbations of the same magnitude | NPE · Six benchmark tasks | KL divergence between clean and perturbed posterior estimates | Hurt |
| Add the moving-average trace of the FIM as a regularizer | NPE · Six benchmark tasks | KL divergence between clean and perturbed posterior estimates | Helped |
| Add the moving-average trace of the FIM as a regularizer | NPE · Pyloric network simulator | KL divergence between clean and perturbed posterior estimates | Helped |
| Use FIM regularization | NPE · Six benchmark tasks | expected coverage under adversarial perturbations | Mixed |
| Use intermediate FIM regularization strength to trade some clean-data accuracy for robustness | NPE · Six benchmark tasks | average log-likelihood on unperturbed data | Hurt |
| Replace FIM regularization with adversarial training (ablation) | NPE · Six benchmark tasks | KL divergence between clean and perturbed posterior estimates | No clear effect |
| Replace FIM regularization with TRADES (ablation) | NPE · Six benchmark tasks | KL divergence between clean and perturbed posterior estimates | No clear effect |
| Replace the moving-average trace-of-FIM regularizer with exact-FIM trace or largest-eigenvalue regularization (ablation) | NPE with a Gaussian density estimator on the VAE task · VAE | adversarial robustness | No clear effect |
| Use a posterior ensemble of 10 MAFs instead of a single NPE estimator (ablation) | NPE · Benchmark tasks | adversarial robustness | No clear effect |
| Add random noise augmentation with epsilon = 1.0 (ablation) | NPE · Benchmark tasks | adversarial robustness | No clear effect |
| Use an MMD attack objective instead of a KL-divergence objective (ablation) | Adversarial attacks on NPE · Lotka Volterra and SIR | attack impact | Hurt |
| Use the forward rather than backward KL-divergence attack objective, or vice versa (ablation) | Adversarial attacks on NPE · Benchmark tasks | attack severity | No clear effect |
Closest papers by meaning
- Robust Amortized Simulation-Based Inference via Learned Error Models (NeurIPS 2026 · similarity 0.90)
- Robust Amortized Bayesian Inference with Self-Consistency Losses on Unlabeled Data (ICLR 2026 · similarity 0.88)
- Amortized Simulation-Based Inference in Generalized Bayes via Neural Posterior Estimation (ICML 2026 · similarity 0.88)
- A PAC-Bayes Analysis of Adversarial Robustness (NeurIPS 2021 · similarity 0.87)
- Leveraging Self-Consistency for Data-Efficient Amortized Bayesian Inference (ICML 2024 · similarity 0.87)
- All-in-one simulation-based inference (ICML 2024 · similarity 0.86)
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.