Research map / Adversarial robustness and security / Certified and empirical robustness
Measuring Representational Robustness of Neural Networks Through Shared Invariances
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 |
|---|---|---|---|
| Train both ResNet18 models with adversarial training instead of Vanilla crossentropy | ResNet18 pair with different initializations · CIFAR10 | STIR: 0.605_{\pm 0.013} -> 0.934_{\pm 0.003} | Helped |
| Use adversarial rather than arbitrary IRIs (ablation) | STIR for ResNet18 pairs · CIFAR10 | STIR: 0.934_{\pm 0.003} -> 0.096_{\pm 0.007} | Hurt |
| Measure shared invariance with STIR rather than observational CKA | Vanilla ResNet18 pair · CIFAR10 | shared invariance estimate: 0.967_{\pm 0.000}; 0.605_{\pm 0.013} | Mixed |
| Increase adversarial-training inner-loop iterations from 1 to 10 | ResNet18 · CIFAR10 | robust accuracy: 34.6_{\pm 0.6} -> 51.5_{\pm 0.6} | Helped |
| Add adversarial training to Vanilla models | Representations across layers · CIFAR10 | shared invariance | Helped |
| Compare later rather than initial layers in Vanilla models | Models trained from different initializations · CIFAR10 | shared invariance | Hurt |
| Train models with residual connections rather than non-residual architectures | ResNet18 and ResNet34 versus VGG models · CIFAR10 | shared invariance | Helped |
| Use different adversarially robust training methods rather than the same AT method | ResNet18 models trained with AT, TRADES, and MART · CIFAR10 | shared invariance | Hurt |
| Add training data at successive update timesteps | ResNet18 model updates · CIFAR10 | STIR between successive models | Helped |
Closest papers by meaning
- On 1/n neural representation and robustness (NeurIPS 2020 · similarity 0.88)
- Understanding Robust Learning through the Lens of Representation Similarities (NeurIPS 2022 · similarity 0.87)
- Grounding Representation Similarity Through Statistical Testing (NeurIPS 2021 · similarity 0.87)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.85)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.85)
- Causal Adversarial Perturbations for Individual Fairness and Robustness in Heterogeneous Data Spaces (AAAI 2024 · similarity 0.85)
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.