Research map / Federated learning and privacy / Differential privacy
PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs
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 PCEvolve, an API-assisted DP synthetic image method that replaces Private Evolution's Gaussian-mechanism similarity voting with an adapted exponential mechanism driven by a contrastive filter and similarity calibrator, for few-shot private data.
Problem
Private Evolution (PE) produces DP synthetic images via diffusion APIs but degrades badly when only a few private images per class are available, because Gaussian-mechanism similarity votes are overwhelmed by noise.
Approach
Aggregate private class centers, use a contrastive filter that keeps synthetic images closer to their own class center than any other, calibrate similarity scores to span [0,1] with a single factor tau, and sample prototypical candidates via the exponential mechanism (sensitivity 1) inside the evolution loop.
Key results
- PCEvolve reaches 64.04% on COVIDx, 69.10% on Camelyon17, 50.95% on KVASIR-f and 59.26% on MVTecAD-l, beating the best baseline PE (59.63, 63.66, 48.88, 57.41).
- PCEvolve surpasses the private-only baseline by up to 8.20% on downstream models and keeps 95% of its accuracy at N=50 synthetic shots (0.95% drop vs 4.78% for PE).
- Removing the contrastive filter (u=h only) drops Camelyon17 accuracy to 59.39%, versus 69.10% for the full method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PCEvolve instead of PE | API-assisted DP synthetic data generation · COVIDx | Top-1 accuracy (%): 59.63 -> 64.04 | Helped |
| use PCEvolve instead of PE | API-assisted DP synthetic data generation · Camelyon17 | Top-1 accuracy (%): 63.66 -> 69.10 | Helped |
| use PCEvolve instead of PE | API-assisted DP synthetic data generation · KVASIR-f | Top-1 accuracy (%): 48.88 -> 50.95 | Helped |
| use PCEvolve instead of PE | API-assisted DP synthetic data generation · MVTecAD-l | Top-1 accuracy (%): 57.41 -> 59.26 | Helped |
| add similarity calibrator h to the contrastive filter g (ablation) | PCEvolve · COVIDx | Top-1 accuracy (%): 56.59 -> 64.04 | Helped |
| add similarity calibrator h to the contrastive filter g (ablation) | PCEvolve · Camelyon17 | Top-1 accuracy (%): 66.51 -> 69.10 | Helped |
| add similarity calibrator h to the contrastive filter g (ablation) | PCEvolve · KVASIR-f | Top-1 accuracy (%): 44.61 -> 50.95 | Helped |
| add similarity calibrator h to the contrastive filter g (ablation) | PCEvolve · MVTecAD-l | Top-1 accuracy (%): 55.74 -> 59.26 | Helped |
| replace uncalibrated h' scores with calibrated similarity scores h (ablation) | PCEvolve with contrastive filter g · COVIDx | Top-1 accuracy (%): 56.11 -> 64.04 | Helped |
| replace uncalibrated h' scores with calibrated similarity scores h (ablation) | PCEvolve with contrastive filter g · Camelyon17 | Top-1 accuracy (%): 65.29 -> 69.10 | Helped |
| replace uncalibrated h' scores with calibrated similarity scores h (ablation) | PCEvolve with contrastive filter g · KVASIR-f | Top-1 accuracy (%): 50.67 -> 50.95 | Helped |
| replace uncalibrated h' scores with calibrated similarity scores h (ablation) | PCEvolve with contrastive filter g · MVTecAD-l | Top-1 accuracy (%): 55.56 -> 59.26 | Helped |
Surprising findings
- Directly applying DP noise to private images (DPImg) can hurt performance, sometimes below the initial model.
- Several one-shot API baselines (RF, GCap, LE, B) barely improve or degrade over the initial downstream model.
Limitations
- Only one hyperparameter tau, which is sensitive: accuracy varies strongly across tau (e.g., MVTecAD-l 40.70 at tau=1 vs 59.26 at tau=10).
- Evaluated on four few-shot datasets with 10 shots per class; broader few-shot scenarios left for future exploration.
- Relies on paid or closed generative APIs; OJ (online) requires payment.
Experimental setup
Datasets: COVIDx Camelyon17 KVASIR-f MVTecAD-l
Benchmarks: COVIDx Camelyon17 KVASIR-f MVTecAD-l
Metrics: Top-1 accuracy classification accuracy score
Baselines: B (t2i prompt only) LE (LLaMA-enhanced prompt) RF (random filtering) GCap (LLaVA captions) DPImg PE PE-EM PCEvolve-GM
Builds on: Private Evolution Exponential Mechanism Gaussian Mechanism RG (random generation DP image replicas) IP-Adapter
Closest papers by meaning
- Private Evolution Converges (NeurIPS 2025 · similarity 0.92)
- Differentially Private Synthetic Data via Foundation Model APIs 1: Images (ICLR 2024 · similarity 0.91)
- A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets (AAAI 2025 · similarity 0.89)
- Differentially Private Synthetic Data via Foundation Model APIs 2: Text (ICML 2024 · similarity 0.89)
- Synthesize Privacy-Preserving High-Resolution Images via Private Textual Intermediaries (NeurIPS 2025 · similarity 0.89)
- Contrastive Private Data Synthesis via Weighted Multi-PLM Fusion (ICML 2025 · 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.