When Is Personalized Federated Learning Worth Its Deployment Cost?
Organizations: Computer Engineering, Istanbul Technical University, Istanbul, Türkiye · Computer Engineering, Marmara University, Istanbul, Türkiye
Abstract
Personalized federated learning (PFL) may improve client-level performance, but deeper personalization increases deployment cost. We formulate personalization-depth selection as a utility-based empirical decision problem in which client utility equals Macro-F1 minus a cost-sensitivity weight times deployment cost, over six personalization depths labeled P0 through P5. We compare fixed policies with an offline per-client Oracle reference and report allocation maps over Dirichlet data heterogeneity and cost sensitivity. Although evaluated on this hierarchy, the formulation can accommodate any finite candidate set. On CIFAR-10, CIFAR-100, and a PTB-XL subset, personalization shows empirical, setting-dependent decision regimes rather than a universal benefit: higher cost sensitivity shifts Oracle mass toward the global-only depth P0; low cost sensitivity can favor deeper actions. Patterns are grid observations (three seeds per cell), not theoretical boundaries. Best-uniform is competitive with Oracle assignment (6.5 percent plus or minus 7.8 percent relative regret; 95 percent confidence interval 5.9 to 7.1), whereas random and personalize-all incur much larger regret (189.3 percent plus or minus 367.1 percent and 409.0 percent plus or minus 803.2 percent). Under the same utility, Ditto full-model personalization is substantially suboptimal once cost is counted. The contribution is an offline characterization of when personalization depth is worth its cost, not a new PFL optimizer or deployment-time scheduler.
Figures & tables
| Action | Locally trainable components |
|---|---|
| None; global-only model | |
| Classifier head | |
| Adapter and head | |
| Last encoder block, adapter, head | |
| Last two encoder blocks, adapter, head | |
| Full model personalization |
| Comp. | Meaning | Pre-weight scale | Weight |
|---|---|---|---|
| Params | Trainable parameters at | 0.25 | |
| FLOPs | Local train+eval estimate | 0.20 | |
| Time | Local training wall time (s) | — | 0.25 |
| Comm | Model-delta bytes | 0.15 | |
| Mem | Param+activation proxy (MB) | 0.15 |
| Qualitative reading | Example regime (illustrative only) | |
|---|---|---|
| 0.0 | Accuracy-only | Ignore cost; Macro-F1 maximization |
| 0.1 | Mild cost preference | Quality dominates; cost secondary |
| 0.3 | Balanced default | Default in sparsity tables |
| 0.5–1.0 | Strong cost preference | Edge / battery-constrained settings |
| 2.0 | Cost-dominant | Cost dominates small F1 gains |
| Policy | Regret (%) | Oracle win (%) | 95% CI |
|---|---|---|---|
| Random | [160.6,219.8] | ||
| Best-uniform | [5.9,7.1] | ||
| Global-only | [25.9,30.6] | ||
| Personalize-all | [346.2,472.9] |
| Policy | Regret (%) | 95% CI | Oracle win (%) |
|---|---|---|---|
| Ditto | [380.7,534.7] | 94.0 | |
| Best-uniform | [5.7,7.1] | 80.7 | |
| Global-only | [25.3,30.6] | 81.7 |
| Dataset | Regret (%) | Win (%) | 95% CI |
|---|---|---|---|
| CIFAR-10 | [4.7,8.3] | ||
| CIFAR-100 | [4.4,6.8] | ||
| CIFAR-10 aux. | [5.9,8.4] | ||
| PTB-XL subset | [6.5,7.9] |
| Dataset | P0 ratio | Deep ratio | Unif. gap | P5 gap |
|---|---|---|---|---|
| CIFAR-100 | ||||
| CIFAR-10 aux. | ||||
| PTB-XL | see Table VIII | |||
| Overall | ||||
| Metric | Mean std | 95% CI |
|---|---|---|
| P0 ratio ( ) | [0.481,0.631] | |
| Deep ratio ( ) | [0.079,0.178] | |
| Oracle vs uniform (rel. %) | [8.5,13.6] | |
| Oracle vs P5-all (rel. %) | [82.2,116.0] | |
| Regret vs uniform (all ) | [6.5,7.9] |
| Policy | Sig. cells | Median | Median | Interp. |
|---|---|---|---|---|
| Random | 588/594 (99.0%) | 1.07 | Strong | |
| Best-uniform | 430/594 (72.4%) | 0.0017 | 0.50 | Small–mod. |
| Global-only | 428/594 (72.1%) | 0.0011 | 0.66 | Moderate |
| Personalize-all | 580/594 (97.6%) | 5.62 | Strong (cost) |
| Item | Purpose |
|---|---|
| Utility curve over | Cost sensitivity |
| P0/deep allocation ratios | Regime patterns |
| Best-uniform baseline | Simple deployment test |
| Oracle regret + CIs | Headroom and stability |
| Cost/weight caveat | Bounds interpretation |
Federated Learning
Federated learning trains shared models across data holders without requiring them to pool their raw data.
See all 1,037 papers