Research map / Adversarial robustness and security / Certified and empirical robustness
Shedding More Light on Robust Classifiers under the lens of Energy-based Models
Mujtaba Hussain Mirza, Maria Rosaria Briglia, Senad Beadini, Iacopo Masi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use KL divergence instead of CE as inner attack loss without weighting (ablation) | TRADES-style adversarial training · CIFAR-100 (ResNet-18) | AutoAttack robust accuracy: 23.97 -> 24.59 | Improved |
| weight entire loss with PM_adv (MAIL-TRADES style) instead of no weighting (ablation) | TRADES on CIFAR-100 · CIFAR-100 (ResNet-18) | AutoAttack robust accuracy: 24.29 -> 23.26 | Hurt |
| add marginal-energy weighting (WEAT) to TRADES (ablation) | TRADES · CIFAR-100 (ResNet-18) | AutoAttack robust accuracy: 24.29 -> 24.88 | Improved |
| use BCE instead of CE in MART outer loss (ablation) | MART-weighted loss · CIFAR-100 (ResNet-18) | PGD vs AutoAttack: PGD 27.32 -> 28.24; AA 23.71 -> 23.63 | Mixed |
| weight loss by energy-based scalar instead of PM weighting (WEAT on CIFAR-10) | TRADES · CIFAR-10 (ResNet-18) | AutoAttack robust accuracy: 49.46 -> 49.75 (WEAT_adv); WEAT_nat 49.02 | No clear effect |
| remove high-energy correctly classified samples from AT (ablation) | SAT robust model · CIFAR-10 | robust error under PGD | Hurt |
| start SGLD from Gaussian per class instead of random noise (ablation) | robust classifier SGLD sampling (l_inf) · CIFAR-10 generation | FID: 357.21 -> 64.98 | Improved |
| use PCA-based SGLD init with momentum instead of Gaussian-per-class init (l_inf) (ablation) | Better DM classifier · CIFAR-10 generation | IS / FID: IS 7.18 -> 7.66; FID 64.98 -> 97.38 | Mixed |
| use PCA init instead of Gaussian init under l_2 threat model (ablation) | Better DM classifier · CIFAR-10 generation | IS / FID: IS 8.75 -> 8.97; FID 27.71 -> 30.74 | Mixed |
| adopt l_2 instead of l_inf robust classifier for generation (ablation) | SGLD generation · CIFAR-10 generation | FID: 97.38 -> 30.74 | Improved |
| train with SAT instead of TRADES (generative quality) | robust classifier generation · CIFAR-10 generation | IS: 7.19 -> 7.96 | Improved |
Closest papers by meaning
- Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation (ICLR 2026 · similarity 0.92)
- Your Classifier Can Do More: Towards Balancing the Gaps in Classification, Robustness, and Generation (CVPR 2026 · similarity 0.90)
- Towards Understanding the Generative Capability of Adversarially Robust Classifiers (ICCV 2021 · similarity 0.90)
- A Unified Contrastive Energy-based Model for Understanding the Generative Ability of Adversarial Training (ICLR 2022 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- Improving Adversarial Energy-Based Model via Diffusion Process (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.
