GAP-FL is a research toolchain for studying the privacy-utility-fairness trade-off in federated learning under cross-silo settings with demographic heterogeneity. In these settings, differential privacy can degrade minority-group performance more strongly than majority-group performance. The main contribution is a group-aware federated architecture in which a shared representation is trained collaboratively under DP-SGD, while group-specific local heads stay on each client and are trained without added noise.
- Code and toolbox link: https://github.com/vtsouval/GAP-FL/
- The repository contains the full training/evaluation pipeline.
- The tool can be used by cloning the repository and installing the Python dependencies listed in
requirements.txt. More detailed instructions are provided in the link.
GAP-FL addresses a practical problem in privacy-preserving federated learning: DP-SGD protects training data, but its injected noise often harms under-represented demographic groups more strongly than others. This is especially problematic in applications where decision-making must be fair, reliable, and robust across diverse populations. The tool combines a new model architecture to study this effect and to evaluate whether group-aware personalization can mitigate it without relaxing standard sample-level differential privacy. The proposed architectural decomposition improves F1 by up to 11.4 points and reduces cross-group disparity by up to 38.2 percent in EO gap under matched privacy budgets.
Key features:
- Introduce and evaluate a group-aware federated architecture that keeps the shared encoder under DP-SGD while letting each demographic group learn its own local classifier head.
- Can be used to enable to control group-specific privacy budgets in FL, giving the ability to control the privacy according to various sensitive attributes (e.g. demographic characteristics).
- Provides a broad fairness-oriented evaluation of the privacy-utility-fairness trade-off in FL in tabular classification tasks.
The tool operates in the following stages:
- Select an experiment configuration, including dataset, heterogeneity scenario, method (
gapfl, fedavg, or fedper), DP noise level, clipping mode, seed, number of clients, and batch size.
- Load a local raw dataset and construct demographic groups from protected attributes such as age, sex, and race, depending on the dataset.
- Build controlled heterogeneity scenarios from predefined bucket mappings and split those buckets into runtime clients using deterministic, validated allocation logic.
- Train the federated model for multiple rounds. In GAP-FL, the shared encoder is updated under DP-SGD, while local group heads are updated separately without injected DP noise.
- Track privacy throughout training using RDP accountants. For GAP-FL, group-level privacy can be inspected while still reporting a conservative client-level guarantee based on the worst group-level epsilon.
- Evaluate each run with utility and fairness metrics across groups, including dedicated summaries for under-represented groups defined from the bottom quartile of pre-split group sizes.
The entrypoints is either ./runb.sh (interactive execution) or main.py (config-driven execution)
This tool is intended for:
- benchmarking FL methods under DP and demographic heterogeneity
- studying how privacy noise affects minority-group performance in high-stakes ML settings
- evaluating architectural alternatives to improve fairness without modifying the underlying DP mechanism
- comparing standard and personalized FL baselines under matched privacy constraints
Within the AIMS5.0 context, GAP-FL is especially relevant for privacy-sensitive multi-institution AI settings, such as healthcare, finance, or public-sector analytics, where FL is attractive but unequal degradation across demographic groups is a serious deployment risk. The tool provides a concrete framework for analyzing that risk and for testing a regulation-aligned mitigation strategy based on architecture rather than privacy relaxation.
- The current repository is a research prototype and experiment framework, not a production-ready FL platform.
- The workflow is primarily CLI- and config-driven; no end-user application with a graphical interface.
- GAP-FL improves the trade-off empirically, but it does not eliminate all fairness disparities, especially under severe heterogeneity or strict privacy budgets.
As this research is still in progress, a link will be provided once the paper is published.