Research map / Adversarial robustness and security / Certified and empirical robustness
Implicit Euler Skip Connections: Enhancing Adversarial Robustness via Numerical Stability
Mingjie Li, Lingshen He, Zhouchen Lin
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 |
|---|---|---|---|
| Replace vanilla skip connections with IE-Skips (K=1 inner iteration) | ResNet-18 naturally trained on MNIST · MNIST FGSM δ=0.15 | Robust Acc (%): +2.91 (88.45 to 91.36) | Improved |
| Increase IE-Skips inner iterations from K=1 to K=3 (ablation) | IE-ResNet-18 on MNIST · MNIST FGSM δ=0.15 | Robust Acc (%): +0.91 (91.36 to 92.27) | Improved |
| Replace vanilla skip connections with IE-Skips (K=1) | ResNet-50 naturally trained on CIFAR-10 · CIFAR-10 FGSM δ=1/255 | Robust Acc (%): +0.80 (67.89 to 68.69) | Improved |
| Replace vanilla skip connections with IE-Skips (K=1) | PreAct ResNet-18/34/50 with PGD-10 adversarial training on CIFAR-10 · CIFAR-10 PGD-20 | Robust Acc (%): +1.02 (ResNet-18), +1.22 (ResNet-34), +1.24 (ResNet-50) | Improved |
| Replace vanilla skip connections with IE-Skips (K=1, gamma=0.1) on WideResNet-34 with TRADES | WideResNet-34-10 TRADES training on CIFAR-10 · CIFAR-10 PGD-20, PGD-100, C&W | Robust Acc (%): +2.56 PGD-20, +3.06 PGD-100, +4.63 C&W vs own reproduction | Improved |
| Replace vanilla skip connections with IE-Skips | ResNet-34 with PGD-3 training on CIFAR-10 · CIFAR-10 C&W | Robust Acc (%): +0.55 (49.70 to 50.25) | Improved |
| Weight-sharing of adjacent residual blocks (SwResNet-58) instead of IE-Skips | ResNet-34 with PGD-3 training on CIFAR-10 · CIFAR-10 C&W and PGD-20 | Robust Acc (%): -13.73 C&W, -0.09 PGD-20 vs ResNet-34 (SwResNet-58 is deeper) | Hurt |
| Increase inner iterations K of IE-Skips from 0 to 5 (ablation) | PreAct ResNet-18 (width 64) with PGD-3 training · CIFAR-10 PGD-20 | Robust Acc (%): robust accuracy rises from roughly 38 to 42 (figure only) | Improved |
| Increase inner iterations K of IE-Skips (ablation) | PreAct ResNet-18 with PGD-3 training · CIFAR-10 natural accuracy | Natural Acc (%) | No clear effect |
Closest papers by meaning
- Robust Neural Networks inspired by Strong Stability Preserving Runge-Kutta methods (ECCV 2020 · similarity 0.89)
- Robust Design of Deep Neural Networks Against Adversarial Attacks Based on Lyapunov Theory (CVPR 2020 · similarity 0.88)
- On Robustness of Neural Ordinary Differential Equations (ICLR 2020 · similarity 0.87)
- Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets (ICLR 2020 · similarity 0.87)
- Can Implicit Bias Imply Adversarial Robustness? (ICML 2024 · similarity 0.87)
- Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation (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.
