How does FedSAP split a neural network across devices that do not trust each other's data?

A new federated learning framework called FedSAP treats structured pruning as a three-way channel allocation problem. It assigns each channel to a global pool, a private pool, or a dropped state so that devices with different compute budgets and different data domains can train together without mixing incompatible features.

Why uniform compression fails on mixed fleets

Federated learning on edge devices has always faced two constraints at once. Devices differ in memory, compute, and battery. Their local data also differ — photos taken indoors versus outdoors, handwritten digits from different countries, office objects captured under different lighting.

Existing resource-adaptive methods decide how much of a model each client trains. They do not decide where the retained capacity should live or how it should be shared. Domain-generalization methods assume every client runs the full architecture. Uniform compression therefore discards high-utility channels. A single aggregation path mixes transferable features with domain-sensitive updates.

Tri-state channel allocation

FedSAP reframes the problem. Each client receives a keep ratio — a single number that represents its resource budget. The framework converts that ratio into non-uniform layer budgets. Every channel in every layer then falls into one of three states. Stable channels go to a Global pool. Useful but domain-sensitive channels go to pseudo-domain-specific Private pools. Low-utility channels go to a Dropped state.

The partition lets broadly useful features benefit from cross-client pooling while isolating domain-sensitive updates from incompatible clients. No channel is shared outside its intended scope.

Domain-Guided Assignment and Type-Matched Aggregation

Pseudo-domains are not given in advance. FedSAP infers them from shallow-gradient similarity across clients. Channels that behave similarly across a subset of devices form a pseudo-domain. Each pseudo-domain receives its own Private pool.

During aggregation, Type-Matched Aggregation restricts each channel to its pool. Global channels average across all clients. Private channels average only inside their pseudo-domain. Dropped channels are never transmitted. The paper does not specify how many pseudo-domains emerge in practice; it only describes the inference mechanism.

Accuracy gains under heavy pruning

Experiments run on two standard benchmarks: Digits and Office-Caltech. Across three random seeds FedSAP reaches 76.00% mean global accuracy on Digits and 72.67% on Office-Caltech. The strongest baseline trails by 1.70 and 4.92 percentage points respectively. Clients operate with pruning ratios up to 80%.

The source does not report per-client latency, energy draw, or communication volume. It also does not describe how the keep ratio is chosen for a given device in a real deployment.

What the preprint leaves open

The method assumes gradient similarity is a reliable proxy for domain membership. No ablation tests that assumption against label-based or metadata-based grouping. The three seeds provide variance estimates but the paper does not report confidence intervals.

Code and data links appear in the arXiv entry but the preprint itself does not guarantee reproducibility artifacts. Deployment on actual heterogeneous hardware — phones, sensors, microcontrollers — remains untested.