Research map / Adversarial robustness and security / Certified and empirical robustness
Triple Wins: Boosting Accuracy, Robustness and Efficiency Together by Enabling Input-Adaptive Inference
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 adaptive early exits and Max-Average adversarial training | SmallCNN · MNIST | ATA (Worst-Case): 96.31% -> 96.43% | Helped |
| Add adaptive early exits and Max-Average adversarial training | SmallCNN · MNIST | TA: 99.49% -> 99.52% | Helped |
| Add adaptive early exits and Max-Average adversarial training | ResNet-38 · CIFAR-10 | ATA (Worst-Case): 42.29% -> 43.32% | Helped |
| Add adaptive early exits and Max-Average adversarial training | ResNet-38 · CIFAR-10 | Average MFlops: 79.42 -> 57.81 | Helped |
| Add adaptive early exits and Max-Average adversarial training | MobileNetV2 · CIFAR-10 | ATA (Worst-Case): 46.92% -> 46.99% | Helped |
| Add adaptive early exits and Max-Average adversarial training | ResNet-38 · CIFAR-10, FGSM | ATA (Worst-Case): 51.11% -> 51.05% | No clear effect |
| Add adaptive early exits and Max-Average adversarial training | ResNet-38 · CIFAR-10, WRM | ATA (Worst-Case): 83.35% -> 83.77% | Helped |
| Use Max-Average rather than Average attack generation (ablation) | RDI-ResNet38 · CIFAR-10 | ATA (Worst-Case): 40.53% -> 43.32% | Helped |
| Use Max-Average rather than Average attack generation (ablation) | RDI-MobileNetV2 · CIFAR-10 | ATA (Worst-Case): 45.93% -> 46.99% | Helped |
| Use Max-Average rather than Average attack generation (ablation) | RDI-SmallCNN · MNIST | ATA (Worst-Case): 96.30% -> 96.43% | No clear effect |
| Train with a single attack from the main branch rather than Max-Average (ablation) | RDI-ResNet38 · CIFAR-10 | ATA (Worst-Case): 43.32% -> 5.58% | Hurt |
| Replace adversarial training with standard training (ablation) | RDI-SmallCNN · MNIST | ATA (Worst-Case): 96.43% -> 1.32% | Hurt |
Closest papers by meaning
- Adaptive Neural Networks for Efficient Inference (ICML 2017 · similarity 0.88)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.87)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.86)
- Accelerating Adversarially Robust Model Selection for Deep Neural Networks via Racing (AAAI 2024 · similarity 0.86)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.86)
- Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks (ICLR 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.