Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MPCPA: Multi-Center Privacy Computing with Predictions Aggregation based on Denoising Diffusion Probabilistic Model

Guibo Luo, Hanwen Zhang, Xiuling Wang, Mingzhi Chen, Yuesheng Zhu

TL;DR

The paper addresses privacy-preserving multi-center learning under Non-IID data and stringent communication constraints. It introduces MPCPA, a framework that combines conditional diffusion model training, DDPM-based data generation, per-client classifiers, and ensemble aggregation of predictions to avoid sharing raw data. Empirical results show MPCPA matches or surpasses centralized learning on several natural and medical datasets and consistently outperforms Federated Learning, while requiring far fewer communications ($3n$ transmissions versus $2n\cdot\text{iters}$). The approach also demonstrates robustness to image memorization and membership inference attacks, suggesting practical applicability for privacy-preserving collaboration in distributed settings.

Abstract

Privacy-preserving computing is crucial for multi-center machine learning in many applications such as healthcare and finance. In this paper a Multi-center Privacy Computing framework with Predictions Aggregation (MPCPA) based on denoising diffusion probabilistic model (DDPM) is proposed, in which conditional diffusion model training, DDPM data generation, a classifier, and strategy of prediction aggregation are included. Compared to federated learning, this framework necessitates fewer communications and leverages high-quality generated data to support robust privacy computing. Experimental validation across multiple datasets demonstrates that the proposed framework outperforms classic federated learning and approaches the performance of centralized learning with original data. Moreover, our approach demonstrates robust security, effectively addressing challenges such as image memorization and membership inference attacks. Our experiments underscore the efficacy of the proposed framework in the realm of privacy computing, with the code set to be released soon.

MPCPA: Multi-Center Privacy Computing with Predictions Aggregation based on Denoising Diffusion Probabilistic Model

TL;DR

The paper addresses privacy-preserving multi-center learning under Non-IID data and stringent communication constraints. It introduces MPCPA, a framework that combines conditional diffusion model training, DDPM-based data generation, per-client classifiers, and ensemble aggregation of predictions to avoid sharing raw data. Empirical results show MPCPA matches or surpasses centralized learning on several natural and medical datasets and consistently outperforms Federated Learning, while requiring far fewer communications ( transmissions versus ). The approach also demonstrates robustness to image memorization and membership inference attacks, suggesting practical applicability for privacy-preserving collaboration in distributed settings.

Abstract

Privacy-preserving computing is crucial for multi-center machine learning in many applications such as healthcare and finance. In this paper a Multi-center Privacy Computing framework with Predictions Aggregation (MPCPA) based on denoising diffusion probabilistic model (DDPM) is proposed, in which conditional diffusion model training, DDPM data generation, a classifier, and strategy of prediction aggregation are included. Compared to federated learning, this framework necessitates fewer communications and leverages high-quality generated data to support robust privacy computing. Experimental validation across multiple datasets demonstrates that the proposed framework outperforms classic federated learning and approaches the performance of centralized learning with original data. Moreover, our approach demonstrates robust security, effectively addressing challenges such as image memorization and membership inference attacks. Our experiments underscore the efficacy of the proposed framework in the realm of privacy computing, with the code set to be released soon.
Paper Structure (26 sections, 7 equations, 9 figures, 7 tables, 1 algorithm)

This paper contains 26 sections, 7 equations, 9 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Federated Learning
  4. Learning From Synthetic Data
  5. Method
  6. DDPM Generation Module
  7. Prediction Aggregation
  8. Communications of MPCPA
  9. Algorithms
  10. Experiments
  11. Dataset
  12. Classification Results
  13. Non-IID Dataset Evaluation and External Testing
  14. Communications
  15. Privacy Analysis
  16. ...and 11 more sections

Figures (9)

  • Figure 1: The proposed multi-center privacy computing framework, which consists of conditional diffusion model training, DDPM data generation, classifiers, and ensemble learning modules
  • Figure 2: Communications between clients and central server.
  • Figure 3: The image distributions of the three Tuberculosis clients.
  • Figure 4: The relationship between performance improvements and training set size.
  • Figure 5: The impact of the number of generated data on accuracy across the four datasets.
  • ...and 4 more figures