This repository contains the implementation, experimental artifacts, figures, and analysis for a research study on client-specific personalization depth in federated learning.
The central question is:
How much of a shared model should each client personalize after federated training?
Personalization depth is not globally optimal. Different clients can prefer different adaptation depths within the same federated run, and forcing all clients to use a single policy leaves measurable performance unrealized.
Personalized federated learning is commonly motivated by client heterogeneity, but many systems still apply the same adaptation strategy to every client. This project studies personalization depth as a client-dependent decision.
The study compares three post-hoc personalization policies:
- Head-only fine-tuning
- Partial fine-tuning
- Full fine-tuning
Experiments are conducted across five benchmark datasets and multiple Dirichlet non-IID regimes.
- Formulates client-specific personalization depth as a distinct federated learning decision problem.
- Evaluates head-only, partial, and full fine-tuning across multiple datasets and heterogeneity regimes.
- Quantifies oracle client-wise routing headroom beyond the strongest fixed policy.
- Analyzes client-level variation in preferred adaptation depth.
- Evaluates lightweight automatic selectors and shows that recovering oracle gains is nontrivial.
- CIFAR-10
- CIFAR-100
- SVHN
- FashionMNIST
- EMNIST-Balanced
Each dataset is evaluated under Dirichlet non-IID client partitions with multiple heterogeneity settings and random seeds.
federated-personalization-depth/
├── src/ # Core implementation
├── configs/ # Dataset-level experiment configurations
├── scripts/ # Reproduction and analysis helpers
├── results/ # Raw and aggregated experiment outputs
├── figures/ # Consolidated final figures
├── analysis/ # Analysis notes and diagnostics
├── docs/ # Methodology and reproducibility documentation
├── notebooks/ # Exploratory notebooks
├── pyproject.toml
├── requirements.txt
└── README.md
Install dependencies:
pip install -r requirements.txtAggregate stored results:
python scripts/aggregate_results.pyVerify figure availability:
python scripts/reproduce_figures.pyThe repository includes stored outputs and final figures so that the main empirical claims can be inspected without rerunning the full experimental sweep.
- Oracle routing is an upper bound, not a deployable method.
- Lightweight selectors do not consistently recover oracle-level gains.
- Experiments focus on controlled benchmark datasets with synthetic Dirichlet partitions.
- Additional validation on real-world federated datasets would strengthen external validity.
This project identifies personalization depth as a distinct axis of heterogeneity in federated learning. It shows that fixed personalization policies leave measurable performance unrealized and that adaptive depth selection is a nontrivial client-level decision problem.
Victor Obarafor
MIT License