FedSAP: Federated Learning with Structured Adaptive Partitioning for Multi-Domain Heterogeneous Edge Devices
Federated learning on edge devices must simultaneously handle varying resource budgets and domain-shifted local data. FedSAP frames this coupled challenge as a structured pruning problem where channels are assigned to Global, Private, or Dropped states. By inferring pseudo-domains from shallow gradient similarity and allocating layer-wise budgets adaptively, the method improves global mean accuracy under aggressive heterogeneous compression.
The paper presents a conceptually clean solution to an important coupled-heterogeneity problem. The tri-state channel formulation is intuitive and the ablations suggest complementary benefits from Domain-Guided Assignment and Selective Global Protection. However, the empirical gains are modest on Digits ($1.70$ points over the strongest baseline) and the evaluation protocol---reporting best-so-far per-domain accuracies that may occur at different rounds---complicates interpretation of the headline numbers.
The strongest aspect is the explicit joint formulation of layer budget and sharing scope. The tri-state channel ontology---Global, Private, Dropped---directly addresses the limitation that compression ratios alone do not specify where capacity should reside or with whom each retained feature should be shared. Ablations in Appendix D demonstrate that both adaptive layer budgets and type-matched aggregation contribute meaningfully, with the full system reaching $76.00\%$ versus $73.5\%$ for a uniform single-path baseline.
Several concerns limit the strength of the conclusions. First, the headline global accuracy is a best-so-far unweighted mean evaluated at a single round, while per-domain columns are each best-so-far and may come from distinct rounds; this decoupled reporting makes the $76.00\%$ figure difficult to interpret as a simultaneous achievement. Second, pseudo-domain discovery relies solely on first-layer gradient cosine similarity, a heuristic the authors acknowledge is conditional on low-level appearance shifts and can fail for semantic domains; with only $10$ clients and $4$ domains, the grouping signal is fragile. Third, privacy risks from transmitting shallow gradients and activation statistics are noted but not experimentally mitigated or audited, leaving practical deployability unclear.
The evidence supports the claim that structured allocation improves global mean accuracy under heterogeneous pruning, but comparisons are narrow and the evaluation metric is lenient. FedSAP is compared against the most relevant baseline, DapperFL, on only two small-scale vision benchmarks with full client participation. The $+4.92$ point gain on Office-Caltech is offset by explicit trade-offs: SVHN drops $5.94$ points and Webcam drops $3.45$ points relative to DapperFL. Because the method optimizes for mean accuracy rather than worst-domain performance, practitioners seeking robust generalization may find the trade-offs unfavorable.
Reproducibility is partially addressed but incomplete. The main hyperparameters---$g_{\mathrm{target}}=0.9$, $\tau_{\mathrm{var}}=0.05$, $\beta_{\mathrm{EMA}}=0.5$---are disclosed, and Appendix C details optimizer settings, batch size, and DGA protocol. However, no public code repository or dataset preprocessing scripts are referenced, and the method involves interdependent algorithmic choices (dynamic masks, EMA smoothing, softmax temperature, clipping bounds) whose exact implementations are critical for reproduction. The three-seed average is acceptable, but the small benchmark scale and absence of standard deviation reporting obscure result stability.
Federated learning (FL) on heterogeneous edge devices must jointly accommodate unequal resource budgets and domain-shifted local data. Existing resource-adaptive methods decide how much of a model each client trains but not where retained capacity should reside or how it should be shared, whereas federated domain-generalization methods usually assume a shared full architecture. Uniform compression can therefore discard high-utility channels, and a single aggregation path can mix transferable features with domain-sensitive updates. We propose FedSAP, a domain-aware heterogeneous FL framework that casts structured pruning as budget-constrained tri-state channel allocation. FedSAP converts each keep ratio into non-uniform layer budgets, assigns stable channels to a Global pool, useful domain-sensitive channels to pseudo-domain-specific Private pools, and low-utility channels to a Dropped state. This partition lets broadly useful features benefit from cross-client pooling while isolating domain-sensitive updates from incompatible clients. Domain-Guided Assignment infers pseudo-domains from shallow-gradient similarity, while Type-Matched Aggregation restricts each channel to its intended sharing scope. Across three random seeds, FedSAP reaches 76.00% and 72.67% mean global accuracy on Digits and Office-Caltech, exceeding the strongest baseline by 1.70 and 4.92 percentage points while supporting client pruning ratios of up to 80% across heterogeneous clients.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.