Research map / Adversarial robustness and security / Certified and empirical robustness
Scalable Verified Training for Provably Robust Image Classification
Sven Gowal, Krishnamurthy (Dj) Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Mann, Pushmeet Kohli
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 IBP verified training instead of Wong et al. at large radius | IBP · MNIST (eps=0.3) | verified error rate: 26.92% -> 8.05% | Improved |
| use IBP instead of Wong et al. | IBP · CIFAR-10 (eps=8/255) | verified error rate: 79.21% -> 67.96% | Improved |
| use IBP instead of Wong et al. at small radius | IBP · CIFAR-10 (eps=2/255) | verified error rate: 49.96% -> 49.98% | No clear effect |
| use IBP instead of Madry et al. adversarial training | IBP · CIFAR-10 (eps=8/255) | PGD error rate: 75.95% -> 65.23% | Improved |
| use IBP instead of Madry et al. adversarial training | IBP · MNIST (eps=0.3) | PGD error rate: 3.73% -> 6.12% | Hurt |
| use IBP instead of Madry et al. at small CIFAR radius | IBP · CIFAR-10 (eps=2/255) | PGD error rate: 42.01% -> 45.09% | Hurt |
| use IBP instead of Madry et al. adversarial training | IBP · CIFAR-10 (eps=8/255) | test error: 20.33% -> 50.51% | Hurt |
| use IBP instead of Wong et al. for training speed | IBP · MNIST small model | training time per epoch: 2 minutes -> 3.5 seconds | Improved |
| train with IBP loss on downscaled ImageNet instead of nominal training | IBP · Downscaled ImageNet (eps=1/255) | PGD error rate: 100.00% -> 90.88% | Improved |
| train with IBP so the IBP bound itself is a tight estimate | IBP · MNIST (eps=0.3) | IBP-only vs MIP verified error: 8.21% -> 8.05% | Improved |
| use cross-entropy instead of hinge or softplus loss on each specification (ablation) | IBP · MNIST, CIFAR-10, SVHN | verified accuracy | Improved |
| elide the last linear layer with the specification (ablation) | IBP · MNIST, CIFAR-10, SVHN | verified accuracy | Improved |
Closest papers by meaning
- Understanding Certified Training with Interval Bound Propagation (ICLR 2024 · similarity 0.93)
- On the Convergence of Certified Robust Training with Interval Bound Propagation (ICLR 2022 · similarity 0.91)
- Towards Stable and Efficient Training of Verifiably Robust Neural Networks (ICLR 2020 · similarity 0.91)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.87)
- Towards General Robustness Verification of MaxPool-based Convolutional Neural Networks via Tightening Linear Approximation (CVPR 2024 · similarity 0.86)
- Tightening Robustness Verification of MaxPool-based Neural Networks via Minimizing the Over-Approximation Zone (CVPR 2025 · 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.
