Research map / Federated learning and privacy / Communication-efficient FL

PRISM: Privacy-Preserving Improved Stochastic Masking for Federated Generative Models

ICLR 2025 · Federated learning and privacy › Communication-efficient FL · approach: Communication-efficient heterogeneous FL · 0 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 PRISM, a federated generative framework that searches for a stochastic binary mask over a fixed random network (strong lottery ticket) trained with MMD loss and a mask-aware dynamic moving average aggregation, cutting communication and final model size.

Problem

Federated training of generative models suffers from high communication cost, unstable GAN training under non-IID data, limited evaluation to simple datasets, and privacy concerns requiring differential privacy.

Approach

Instead of updating weights, clients learn scores over a frozen random Kaiming-initialized generator, sample Bernoulli binary masks as uplink messages, train with MMD loss on VGG features, and the server aggregates masks with a moving average whose coefficient is set by the Hamming distance between consecutive global masks (MADA).

Key results

• Under IID with (9.8, 1e-5)-DP, PRISM reaches FID 27.30 on MNIST, 46.17 on FMNIST and 48.998 on CelebA, versus GS-WGAN's 71.10, 119.26, 230.79 and DP-FedAvgGAN's 111.09, 118.51, 221.34.
• PRISM uses about 5.75MB communication per round and 7.25MB storage, compared to 14-15MB for GAN baselines.
• Without DP in non-IID, PRISM* (hybrid) reaches FID 6.96 on MNIST and 13.02 on CelebA, outperforming MD-GAN (37.80 and 18.91 on those datasets).

Measured effects reported in the paper

Surprising findings

• Multi-FLGAN diverges under Dirichlet non-IID partitioning without DP.

Limitations

• Non-IID CelebA split is weak heterogeneity (gender-based), and Dirichlet results are mixed for FMNIST and CIFAR10.
• Hybrid PRISM* requires sending float scores for some layers, increasing communication cost.
• Hyperparameters were not tuned; performance may improve with tuning.

Experimental setup

Datasets: MNIST FMNIST CelebA CIFAR10

Benchmarks: MNIST FMNIST CelebA CIFAR10

Metrics: FID precision & recall density & coverage communication cost (MB) storage (MB)

Baselines: MD-GAN Multi-FLGAN DP-FedAvgGAN GS-WGAN

Builds on: edge-popup FedPM GFMN strong lottery ticket hypothesis Kaiming normal initialization

Closest papers by meaning

• Leveraging Randomness in Model and Data Partitioning for Privacy Amplification (ICML 2025 · similarity 0.86)
• Sketched Gaussian Mechanism for Private Federated Learning (NeurIPS 2025 · similarity 0.86)
• Beyond Local Sharpness: Communication-Efficient Global Sharpness-aware Minimization for Federated Learning (CVPR 2025 · similarity 0.86)
• FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.86)
• FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.86)
• Federated Learning via Input-Output Collaborative Distillation (AAAI 2024 · similarity 0.86)