Federated SHAP: Privacy-Preserving and Consistent Post-hoc Explainability in Federated Learning

The widespread adoption of Artificial Intelligence in everyday activities highlights a growing and urgent need for trustworthiness. Designing trustworthy AI systems requires addressing key technical challenges, including ensuring data privacy and model explainability. Federated Learning (FL) is a widely adopted paradigm to preserve data privacy in collaborative learning scenarios, while post-hoc methods are commonly applied to enhance the explainability of opaque AI-based models. In this paper, we propose a novel approach, called Federated SHAP, to simultaneously address privacy and explainability. Specifically, we leverage the SHapley Additive exPlanations (SHAP) method to provide post-hoc explanations of Neural Networks trained through FL. SHAP relies on a representative background dataset; however, constructing such a dataset in the FL setting is particularly challenging since raw data distributed across multiple clients cannot be shared directly due to strict privacy requirements. To address this challenge, we propose two tailored strategies depending on the data type: for tabular data, we adopt a Federated Fuzzy C-Means clustering algorithm to collaboratively summarize the distributed datasets into a suitable background dataset; for image data, we introduce a Federated Generative Adversarial Network (GAN) to synthesize representative background instances. A comprehensive experimental evaluation demonstrates the effectiveness and robustness of our proposed approaches, comparing them against several baseline and alternative strategies in terms of both representativeness and quality of generated explanations. Compared to baselines employing randomly generated representative background datasets, our approach reduces the discrepancy of SHAP explanations by up to three times on tabular data and two times on image data (depending on the test case involved), when measured against the centralized SHAP values computed using the full training set as background dataset.