Research map / Federated learning and privacy / Differential privacy
Reinforcement-Guided Synthetic Data Generation for Privacy-Sensitive Identity Recognition
Xuemei Jia, Jiawei Du, Hui Wei, Jun Chen, Joey Tianyi Zhou, Zheng 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 |
|---|---|---|---|
| add Dynamic Sample Selection (DSS) to Base-DiT (ablation) | Base-DiT · CASIA-WebFace subset face verification | avg accuracy: gain of 2.2% | Improved |
| add semantic reward (SC) (ablation) | Base-DiT + DSS · CASIA-WebFace subset face verification | avg accuracy: 78.18% | Improved |
| add coverage reward (DC) (ablation) | Base-DiT + DSS + SC · CASIA-WebFace subset face verification | avg accuracy: 78.18% -> 78.65% | Improved |
| add expression reward (ED) (ablation) | Base-DiT + DSS + SC + DC · CASIA-WebFace subset face verification | avg accuracy: 78.65% -> 79.07% | Improved |
| reinforcement-guided DiT synthetic augmentation vs baseline | ResNet-50 · Market-1501 | mAP: 85.4 -> 88.6 (3.2% gain) | Improved |
| reinforcement-guided DiT synthetic augmentation vs baseline | ResNet-50 · CUHK03-NP | mAP: 76.6% (2.5% gain) | Improved |
| use text-guided synthetic data from GIF-SD | ResNet-50 · Market-1501 | mAP: mAP decrease of 10.5% vs. baseline | Hurt |
| use simulation-based augmentation (FineGPR, InfinitePerson) | ResNet-50 · Market-1501 | mAP: 82.4 and 57.3 vs 85.4 | Hurt |
| real-image augmentation (R-Erasing, CIDAM, CaAug) | ResNet-50 · Market-1501 | mAP: about 2% mAP gain | Improved |
| our method vs NegFaceDiff | face recognition ResNet-50 + CosFace · CASIA-WebFace subset face verification | avg accuracy: surpassing by 0.94% | Improved |
| our synthetic data vs prior methods on Indian and Asian subsets | face recognition · RFW | accuracy: Indian 1.0%, Asian 0.9% | Improved |
| use ViT backbone with our method | ViT-16 · Market-1501 / CUHK03-NP | rank-1 / mAP: clear Rank-1 boosts, mAP comparable to R-Erasing | Mixed |
Closest papers by meaning
- A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets (AAAI 2025 · similarity 0.88)
- Privacy Amplification Persists under Unlimited Synthetic Data Release (NeurIPS 2026 · similarity 0.88)
- Private Evolution Converges (NeurIPS 2025 · similarity 0.88)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.87)
- PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs (ICML 2025 · similarity 0.87)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.87)
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.
