Research map / Adversarial robustness and security / Attacks on text and graphs
Towards Understanding the Generative Capability of Adversarially Robust Classifiers
Yao Zhu, Jiacheng Ma, Jiacheng Sun, Zewei Chen, Rongxin Jiang, Yaowu Chen, Zhenguo Li
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 |
|---|---|---|---|
| add adversarial training to a normal classifier (lower joint and marginal Fisher scores) | WideResNet-28-10 classifier · CIFAR-10 | Joint Fisher Score: 6.97 -> 2.83 | Improved |
| use E(x,y) only instead of E(x,y)-E(x) when generating images and training adversarial examples (PreJEAT) (ablation) | adversarially robust classifier · CIFAR-10 generation | image quality (qualitative) | Improved |
| add joint likelihood training of p(x,y) with SGLD to PreJEAT (JEAT vs adversarial training) | robust classifier · CIFAR-10 generation | Inception Score: 7.50 -> 8.80 | Improved |
| JEAT instead of adversarial training | robust classifier · CIFAR-10 generation | FID: 60.90 -> 38.24 | Improved |
| JEAT instead of JEM | hybrid generative classifier · CIFAR-10 generation | Inception Score: 8.66 -> 8.80 | Improved |
| JEAT instead of JEM | hybrid generative classifier · CIFAR-10 generation | FID: 38.40 -> 38.24 | No clear effect |
| JEAT instead of JEM | hybrid generative classifier · CIFAR-10 PGD-20 eps=8/255 | robust accuracy (%): 6.11% -> 30.55% | Improved |
| JEAT instead of JEM | hybrid generative classifier · CIFAR-10 | standard accuracy (%): 92.90% -> 85.16% | Hurt |
| JEAT instead of JEM | hybrid generative classifier · CIFAR-10 | Joint Fisher Score: 0.330 -> 0.023 | Improved |
| JEAT instead of JEM | hybrid generative classifier · CIFAR-10 | Marginal Fisher Score: 1.44 -> 0.15 | Improved |
| attack using only E(x,y) instead of full cross-entropy loss (ablation) | normal classifier WRN-28-10 · CIFAR-10 | accuracy under attack (%): 0.12% -> 0.11% | No clear effect |
| inject stronger Gaussian noise during normal classifier training (ablation) | normal classifier · CIFAR-10 generation | image recognizability (qualitative): noise std 8/255 -> 24/255 | Improved |
Closest papers by meaning
- Your Classifier Can Do More: Towards Balancing the Gaps in Classification, Robustness, and Generation (CVPR 2026 · similarity 0.90)
- Shedding More Light on Robust Classifiers under the lens of Energy-based Models (ECCV 2024 · similarity 0.90)
- A Unified Contrastive Energy-based Model for Understanding the Generative Ability of Adversarial Training (ICLR 2022 · similarity 0.90)
- Learning Energy-Based Models with Adversarial Training (ECCV 2022 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- DAT: Improving Adversarial Robustness via Generative Amplitude Mix-up in Frequency Domain (NeurIPS 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.
