Research map / Adversarial robustness and security / Certified and empirical robustness
Shift Invariance Can Reduce Adversarial Robustness
Vasu Singla, Songwei Ge, Ronen Basri, David Jacobs
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 a shift-invariant CNN instead of an FC network | Classifier on orthogonal-vector synthetic data · Orthogonal vectors, n=50, 100, and 200 | average l2 adversarial distance: 0.113 -> 0.033; 0.083 -> 0.029; 0.060 -> 0.024 | Hurt |
| Use a shift-invariant CNN instead of an FC network | Classifier on orthogonal-frequency synthetic data · Orthogonal frequencies, n=50, 100, and 200 | average l2 adversarial distance: 0.129 -> 0.272; 0.126 -> 0.251; 0.103 -> 0.326 | Improved |
| Increase the shared class component p | FC network on dataset 3 · Dataset 3, n=50, 100, and 200 | average l2 adversarial distance: n=50: 0.106 at p=0 -> 0.181 at p=0.3; n=100: 0.080 at p=0 -> 0.181 at p=0.3; n=200: 0.060 at p=0 -> 0.170 at p=0.3 | Improved |
| Increase shift invariance by increasing circular padding | CNN on MNIST · MNIST, PGD l2 and l∞ attacks | robust accuracy | Hurt |
| Replace the shift-invariant CNN with an FC network | Classifier on MNIST and Fashion-MNIST · MNIST and Fashion-MNIST, PGD l2 and l∞ attacks | robust accuracy | Improved |
| Use a ResNet instead of an FC network | Classifier on SVHN · SVHN, PGD l2 and l∞ attacks | robust accuracy | Hurt |
| Use AlexNet or Sparse-MLP instead of ResNet | Classifier on CIFAR-10 · CIFAR-10, adversarial attacks | robust accuracy | Improved |
| Use a ViT instead of a BiT-ResNet | ImageNet classifier · ImageNet-1K, PGD l2 and l∞ attacks | robust accuracy | Improved |
| Increase input dimension while keeping the margin from shrinking | FC network on dataset 3 · Dataset 3 | average l2 adversarial distance | No clear effect |
Closest papers by meaning
- Excessive Invariance Causes Adversarial Vulnerability (ICLR 2019 · similarity 0.89)
- Fundamental Tradeoffs between Invariance and Sensitivity to Adversarial Perturbations (ICML 2020 · similarity 0.88)
- Hold me tight! Influence of discriminative features on deep network boundaries (NeurIPS 2020 · similarity 0.87)
- Blessings and Curses of Covariate Shifts: Adversarial Learning Dynamics, Directional Convergence, and Equilibria (JMLR 2024 · similarity 0.87)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.86)
- Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness (NeurIPS 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.
