Research map / Label-efficient and robust learning / Recognition and long-tail
FairDD: Fair Dataset Distillation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FairDD, a plug-in fairness-aware dataset distillation framework that matches synthetic data to each protected-attribute group synchronously, improving fairness of models trained on condensed datasets across DM and DC.
Problem
Dataset distillation condenses data for accuracy on target attributes but ignores fairness w.r.t. protected attributes; when original data is PA-imbalanced, condensed sets inherit and amplify bias toward majority groups.
Approach
Partition the training set by protected attribute and align synthetic samples synchronously to each PA-group's mean embedding/gradient with equal weight, instead of aligning to the whole class distribution dominated by majority groups. Theory shows the optimum is the unweighted average of group centers and the objective upper-bounds the vanilla objective.
Key results
- On C-MNIST (FG) at IPC=10, DM+FairDD reduces DEO_M from 100.0 to 17.04 and DEO_A from 99.96 to 7.95, while raising accuracy from 25.01 to 94.61.
- DM+FairDD on CIFAR10-S at IPC=10 reaches 45.17 accuracy vs 37.88 for DM, with DEO_M 31.75 vs 59.20.
- FairDD improves fairness across DM, DC, IDC and DREAM on several biased datasets, with DC+FairDD outperforming Whole on some fairness metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FairDD synchronized group-wise matching to DM | DM · C-MNIST (FG), IPC 10 | DEO-M: 100.0 -> 17.04 | Helped |
| Add FairDD synchronized group-wise matching to DM | DM · C-MNIST (FG), IPC 10 | accuracy: 25.01 -> 94.61 | Helped |
| Add FairDD synchronized group-wise matching to DM | DM · CIFAR10-S, IPC 10 | DEO-A: 39.31 -> 8.73 | Helped |
| Replace synchronized alignment with inverse sample-size group weighting (ablation) | DM+FairDD · C-MNIST (FG), IPC 10 | accuracy: 94.61 -> 94.18 | Hurt |
| Replace synchronized alignment with LDAM group weighting (ablation) | DM+FairDD · C-MNIST (FG), IPC 10 | accuracy: 94.61 -> 94.06 | Hurt |
| Replace synchronized alignment with maximum-loss GroupDRO alignment (ablation) | DM+FairDD · C-MNIST (FG), IPC 10 | accuracy: 94.61 -> 92.73 | Hurt |
| Use fairness-aware model-training regularization instead of FairDD (ablation) | DC · CMNIST-BG | DEO-M: 20.66 -> 99.50 | Hurt |
| Use inverse group-proportion weighting during distillation instead of FairDD (ablation) | DC · CMNIST-BG | DEO-M: 20.66 -> 76.44 | Hurt |
| Initialize synthetic samples with a balanced number from each group rather than the compared initializations (ablation) | Vanilla DM · C-MNIST (FG) | DEO-M | No clear effect |
Surprising findings
- Random sampling sometimes yields better fairness than vanilla DDs, e.g., on CelebA.
- Fairness regularization on vanilla DDs (distillation fairness or model training fairness weighting) reduces but does not eliminate the bias.
Limitations
- Requires protected attribute labels for group partitioning; not applicable when PA is unavailable.
- Extra training time grows with the number of groups, especially for gradient matching (DC).
- Theoretical convexity guarantee assumes convex distance; non-convex case only supported empirically.
Experimental setup
Datasets: C-MNIST Colored-FMNIST CIFAR10-S CelebA UTKFace BFFHQ ImageNet Subset-S Tiny-ImageNet-S
Benchmarks: C-MNIST (FG) C-MNIST (BG) C-FMNIST (FG) C-FMNIST (BG) CIFAR10-S CelebA ImageNet Subset-S Tiny-ImageNet-S
Metrics: DEO_M DEO_A accuracy
Baselines: DM DC IDC DREAM Random Whole LW LfF MTT
Builds on: Distribution Matching Dataset Condensation with Gradient Matching IDC DREAM LfF LDAM GroupDRO
Closest papers by meaning
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.87)
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.87)
- FairWASP: Fast and Optimal Fair Wasserstein Pre-processing (AAAI 2024 · similarity 0.86)
- Do Counterfactually Fair Image Classifiers Satisfy Group Fairness? -- A Theoretical and Empirical Study (NeurIPS 2024 · similarity 0.86)
- Fair Deepfake Detectors Can Generalize (NeurIPS 2025 · similarity 0.86)
- FairDeDup: Detecting and Mitigating Vision-Language Fairness Disparities in Semantic Dataset Deduplication (CVPR 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.