Research map / Adversarial robustness and security / Certified and empirical robustness
Learning Adversarially Robust Representations via Worst-Case Mutual Information Maximization
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 |
|---|---|---|---|
| Learn the representation by worst-case mutual information maximization instead of standard mutual-information representation learning | Baseline-H representation with standard MLP · CIFAR-10 | adversarial accuracy: 0.22 \pm 0.08 -> 14.06 \pm 0.69 | Helped |
| Learn the representation by worst-case mutual information maximization instead of standard mutual-information representation learning | Baseline-H representation with standard MLP · CIFAR-10 | natural accuracy: 58.77 \pm 0.22 -> 62.54 \pm 0.12 | Helped |
| Adversarially train the downstream MLP on the worst-case MI representation (ablation) | Worst-case MI representation + standard MLP · CIFAR-10 | adversarial accuracy: 14.06 \pm 0.69 -> 31.52 \pm 0.31 | Helped |
| Adversarially train the downstream MLP on the worst-case MI representation (ablation) | Worst-case MI representation + standard MLP · CIFAR-10 | natural accuracy: 62.54 \pm 0.12 -> 52.34 \pm 0.17 | Hurt |
| Use a worst-case MI representation rather than a standard representation with an adversarially trained MLP (ablation) | Baseline-H with robust MLP · CIFAR-10 | adversarial accuracy: 15.08 \pm 0.63 -> 31.52 \pm 0.31 | Helped |
| Use early stopping based on best adversarial accuracy for the downstream MLP (ablation) | Worst-case MI representation + MLP · CIFAR-10 | adversarial accuracy: 31.52 \pm 0.31 -> 27.53 \pm 0.81 | Hurt |
| Use a worst-case MI representation rather than a standard representation with a standard linear classifier | Baseline-H with linear classifier · CIFAR-10 | adversarial accuracy: 0.15 \pm 0.03 -> 10.98 \pm 0.49 | Helped |
| Adversarially train the downstream linear classifier on the worst-case MI representation (ablation) | Worst-case MI representation + standard linear classifier · CIFAR-10 | adversarial accuracy: 10.98 \pm 0.49 -> 25.15 \pm 0.10 | Helped |
| Adversarially train the downstream linear classifier on the worst-case MI representation (ablation) | Worst-case MI representation + standard linear classifier · CIFAR-10 | natural accuracy: 50.29 \pm 0.58 -> 43.55 \pm 0.10 | Hurt |
| Maximize worst-case mutual information instead of standard mutual information | Unsupervised representation learning for downstream classification · CIFAR-10 | adversarial accuracy | Helped |
Closest papers by meaning
- Adversarial Robustness of Supervised Sparse Coding (NeurIPS 2020 · similarity 0.91)
- Understanding Robust Learning through the Lens of Representation Similarities (NeurIPS 2022 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
- Adversarially Robust Generalization Requires More Data (NeurIPS 2018 · similarity 0.89)
- On Transfer of Adversarial Robustness from Pretraining to Downstream Tasks (NeurIPS 2023 · similarity 0.89)
- Adversarially Robust Multi-task Representation Learning (NeurIPS 2024 · similarity 0.88)
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.