Research map / Federated learning and privacy / Privacy attacks and secure learning

CaPS: Collaborative and Private Synthetic Data Generation from Distributed Sources

ICML 2024 · Federated learning and privacy › Privacy attacks and secure learning · approach: Encryption-based private learning · 1 citations (OpenAlex)

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 CaPS/DP-in-MPC framework that runs select and measure steps of marginal-based DP synthetic data generators (MWEM+PGM, AIM) inside secure multi-party computation over secret-shared distributed data, matching centralized utility while adding input privacy.

Problem

Generating differentially private synthetic tabular data from distributed data holders who cannot share raw data with a trusted aggregator, for arbitrary (horizontal, vertical, mixed) partitionings.

Approach

Secret-share each holder's data and run the DP-relevant select and measure steps of select-measure-generate SDG algorithms inside MPC servers that emulate a trusted curator (DP-in-MPC), publishing only noisy marginals for post-processing generation.

Key results

• On COMPAS, CaPS (horizontal and vertical) matches centralized CDP for AIM: LR-AUC 0.66-0.68 vs 0.67 and error delta 0.015-0.019 vs 0.017.
• Under 3PC passive MPC, the select protocol for AIM costs about 1.19 s and the measure protocol about 0.004 s, with pCOMP at 2.88 s on the Diabetes setup.
• The framework is roughly 1.5-2x slower than CDP in some settings, e.g. AIM on COMPAS takes 301 s vertically vs 153 s centrally.

Measured effects reported in the paper

Surprising findings

• Vertical distribution with AIM was faster than horizontal in runtime because one-way marginals can be computed locally by holders.
• Under 3PC passive, the MPC overhead of the measure step is negligible relative to total runtime.

Limitations

• Utility differences on the small breast-cancer dataset show higher variance, attributed to small data size.
• MPC runtimes are substantially higher than centralized generation, especially for vertical MWEM+PGM and 2PC communication costs.
• Evaluation used a simulated environment on a single machine for the end-to-end runtimes.

Experimental setup

Datasets: breast-cancer COMPAS diabetes

Benchmarks: breast-cancer COMPAS diabetes

Metrics: AUC-ROC F1 workload error runtime (s) communication cost (MB)

Baselines: CDP (centralized DP with trusted aggregator)

Builds on: MWEM+PGM AIM MP-SPDZ replicated secret sharing

Closest papers by meaning

• A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets (AAAI 2025 · similarity 0.88)
• SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.87)
• Private Evolution Converges (NeurIPS 2025 · similarity 0.87)
• Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.86)
• Struct-Bench: A Benchmark for Differentially Private Structured Text Generation (NeurIPS 2025 · similarity 0.86)
• PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs (ICML 2025 · similarity 0.86)