Explainable Federated Learning via Global–Local Attribution Alignment

Federated learning enables on-device training without centralizing data, yet existing systems still struggle to provide explanations that are both locally faithful and globally consistent under strict privacy and bandwidth constraints. Prior approaches either keep explanations siloed across clients, transmit heavy or sensitive artifacts, or replace expressive task models with interpretable surrogates that sacrifice accuracy. We propose xFedAlign, a model-agnostic framework that decouples task optimization in parameter space from explanation coordination in a compact group space. Each client distills a lightweight surrogate to produce private, per-class top-k attribution artifacts, which are robustly aggregated by the server into a Global Explanation Prior that softly aligns client explanations without constraining task learning. Across image, text, and tabular benchmarks with IID and non-IID partitions, xFedAlign matches FedAvg accuracy while consistently reducing explanation drift and improving deletion and insertion AUC relative to Local-XAI, FedAttr-Agg, and Fed-XAI, with only a few kilobytes of additional communication per round. Privacy and robustness evaluations further demonstrate reduced membership inference advantage and increased resistance to attribution poisoning, enabling consistent and trustworthy explanations in federated learning.

When hospitals, banks, or phones train AI models together, they often keep their data private and only share the model itself, a setup called federated learning. But even when these models work well, it is hard to know why they make a given prediction, and the reasons can quietly differ from one hospital or phone to the next, which makes the system harder to trust and audit. We built xFedAlign, a method that lets each participant generate a small, private "explanation summary" of what its model is paying attention to, and then carefully combines these summaries into a shared reference that everyone can agree on. Crucially, the underlying model keeps learning normally, only the explanations are coordinated, and the summaries sent over the network are tiny, noisy, and stripped of identifying detail. Across images, text, and tabular data, including Alzheimer's brain scans and credit card fraud, xFedAlign matched standard federated learning's accuracy while producing far more consistent explanations and resisting attacks that try to either steal training data or manipulate the explanations.