[Baseline] Add SSFL (Salient Sparse Federated Learning)

Type

New Baseline

Description

I would like to contribute a Flower Baseline for SSFL: Discovering Unified Sparse Subnetworks at Initialization for Efficient Federated Learning.

• Paper (TMLR / OpenReview): https://openreview.net/forum?id=kUZ6LhUB26
• Reference implementation: https://github.com/riohib/SSFL

Method

SSFL (Salient Sparse Federated Learning) discovers a unified sparse subnetwork at initialization by aggregating client-local saliency scores. Clients then train within this fixed shared subspace and communicate only active parameters. The paper provides early evidence that learning in a unified sparse subspace can improve the accuracy–sparsity trade-off compared with dynamic pruning methods that continually change the sparse topology during training, while reducing communication relative to dense FL.

Why this is a useful Flower baseline

• Reproduces a published method with a clear, self-contained pipeline .
• Targets a common research need: sparse / communication-efficient FL with non-IID data.
• Fits Flower’s stack well (PyTorch, Flower Datasets, Message API, sparse payload transport).
• Provides paper-matching CIFAR-10/100 configs that others can extend or compare against.

Planned experiments

• CIFAR-10 and CIFAR-100
• ResNet-18
• Static SSFL at density 0.5
• 100 clients, 10% participation per round
• Balanced Dirichlet partitioning with α = 0.3
• Paper-matching optimizer and training configurations for CIFAR-10 and CIFAR-100

Planned Implementation

1. Port the our Flower App into baselines/ssfl.
2. Reproduce SSFL’s saliency-based mask discovery and sparse training using Flower Datasets and the Message API.
3. Add paper configurations and documentation for CIFAR-10/100.

Additional Context

I am one of the paper’s authors.

A working Flower App port and paper-scale runs already exist and we are testing to share soon. We plan to contribute the CIFAR-10/100 static SSFL baseline first, then potentially add more experiments later.

The baseline will use PyTorch, Flower Datasets, and the Flower Message API.