Research map / Label-efficient and robust learning / Knowledge distillation
Fairness without Demographics through Knowledge Distillation
Junyi Chai, Taeuk Jang, Xiaoqian Wang
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 hard labels with softmax soft labels from teacher | ResNet-18 student · COMPAS race | equalized odds: 38.34±3.37% -> 21.32±1.97% | Improved |
| replace hard labels with linear soft labels from teacher | ResNet-18 student · CelebA gender classification (age) | equalized odds: 17.32±1.78% -> 9.64±2.14% | Improved |
| replace hard labels with softmax soft labels | ResNet-18 student · New Adult race | equalized odds: 14.43±1.24% -> 10.34±1.22% | Improved |
| replace hard labels with softmax soft labels | ResNet-18 student · CelebA attractiveness (gender) | accuracy: 81.67±0.63% -> 80.87±0.14% | Hurt |
| replace hard labels with softmax soft labels | ResNet-18 student · COMPAS race | accuracy: 64.13±0.32% -> 63.47±0.44% | Hurt |
| use soft labels (binary to linear) in preliminary label experiment (ablation) | student on new Adult · New Adult race | equalized odds: 13.22±1.70% -> 10.13±1.06% | Improved |
| assign random labels instead of teacher soft labels (ablation) | student on new Adult · New Adult race | equalized odds: 13.22±1.70% -> 13.15±1.12% | No clear effect |
| add DRO (Max-Min fairness) training | ResNet-18 student · CelebA gender classification (age) | accuracy: 90.43±0.23% -> 74.46±0.37% | Hurt |
| add DRO (Max-Min fairness) training | ResNet-18 student · COMPAS race | equalized odds: 38.34±3.37% -> 30.43±3.24% | Mixed |
| add ARL adversarial reweighting | ResNet-18 student · New Adult gender | equalized odds: 18.82±1.82% -> 16.52±1.96% | No clear effect |
| add FairRF | ResNet-18 student · New Adult gender | equalized odds: 18.82±1.82% -> 13.54±1.26% | Improved |
| use large overfitted teacher model directly instead of student | ResNet-152 teacher · COMPAS race | equalized odds: 38.34±3.37% -> 39.64% | Hurt |
Closest papers by meaning
- Towards Harmless Rawlsian Fairness Regardless of Demographic Prior (NeurIPS 2024 · similarity 0.87)
- Fair Dataset Distillation via Cross-Group Barycenter Alignment (ICML 2026 · similarity 0.87)
- Utility-Fairness Trade-Offs and How to Find Them (CVPR 2024 · similarity 0.86)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.85)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.85)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 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.
