Research map / Adversarial robustness and security / Certified and empirical robustness
Improving Robust Generalization by Direct PAC-Bayesian Bound Minimization
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 top-layer TrH regularization to AT | ViT-L16 and Hybrid-L16 · CIFAR-10/100, ℓ∞ (8/255) | AutoAttack accuracy (%): 60.8 -> 61.7; 31.7 -> 32.8; 61.7 -> 61.0; 31.8 -> 32.4 | Mixed |
| Add top-layer TrH regularization to TRADES | ViT-L16 and Hybrid-L16 · CIFAR-10/100, ℓ∞ (8/255) | AutoAttack accuracy (%): 60.3 -> 61.4; 31.4 -> 32.1; 60.8 -> 61.7; 30.6 -> 34.1 | Mixed |
| Add top-layer TrH regularization to AT | ViT-B16 · ImageNet, ℓ∞ (4/255) and ℓ2 (3.0) | AutoAttack accuracy (%): 39.0 -> 42.2; 39.3 -> 42.4 | Helped |
| Add top-layer TrH regularization to AT | ViT-L16 · ImageNet, ℓ∞ (4/255) and ℓ2 (3.0) | AutoAttack accuracy (%): 44.1 -> 48.8; 43.5 -> 47.0 | Helped |
| Add top-layer TrH regularization to TRADES | ViT-B16 · ImageNet, ℓ∞ (4/255) and ℓ2 (3.0) | AutoAttack accuracy (%): 39.3 -> 41.9; 39.7 -> 41.6 | Helped |
| Add top-layer TrH regularization to TRADES | ViT-L16 · ImageNet, ℓ∞ (4/255) and ℓ2 (3.0) | AutoAttack accuracy (%): 40.9 -> 48.9; 41.5 -> 47.3 | Helped |
| Reduce TRADES penalty λ_t from 6 to 1 (ablation) | TRADES(TrH), ViT-L16 · ImageNet | clean and AutoAttack accuracy (%): (70.1, 48.9) -> (78.7, 46.7); (68.9, 47.3) -> (77.1, 45.2) | Mixed |
| Vary the TrH penalty λ (ablation) | AT(TrH) and TRADES(TrH), ViT-L16 · ImageNet, ℓ∞ | AutoAttack accuracy | No clear effect |
| Restrict TrH regularization to the top layer instead of all layers (ablation) | Three-layer network trained with AT · Two Moons | full-network TrH | No clear effect |
| Add TrH regularization to AT | ViT-L16 · CIFAR-10, batch size 32 | peak memory (MB/chip): $5.0 \times10^3$ -> $5.0\times10^3$ | No clear effect |
| Add TrH regularization to AT | ViT-L16 · CIFAR-10, batch size 32 | images per second per chip: 5.5 -> 5.2 | Hurt |
Closest papers by meaning
- PAC-Bayesian Spectrally-Normalized Bounds for Adversarially Robust Generalization (NeurIPS 2023 · similarity 0.92)
- A PAC-Bayes Analysis of Adversarial Robustness (NeurIPS 2021 · similarity 0.90)
- Deterministic PAC-Bayesian generalization bounds for deep networks via generalizing noise-resilience (ICLR 2019 · similarity 0.88)
- PAC-Bayesian Generalization Bounds for Adversarial Generative Models (ICML 2023 · similarity 0.87)
- Provable Robust Overfitting Mitigation in Wasserstein Distributionally Robust Optimization (ICLR 2025 · similarity 0.86)
- Enhancing Robust Fairness via Confusional Spectral Regularization (ICLR 2025 · 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.