FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning

XAI methods can solve practical federated learning problems beyond transparency: The paper demonstrates that explainable AI techniques, traditionally used to help users understand model decisions, can be repurposed to improve collaborative machine learning by identifying which data features are most important for the task at hand.

Selective data sharing can improve collaborative learning while maintaining privacy: FedXDS uses feature attribution to identify only the most task-relevant data elements that need to be shared between organizations, rather than sharing raw data broadly, achieving "higher accuracy and faster convergence compared to existing methods" while providing formal privacy guarantees.

Data heterogeneity remains a key challenge in federated learning: When different organizations have different types of data (statistical heterogeneity), collaborative model training typically performs worse—FedXDS addresses this by intelligently identifying which data should be shared to align contributions across clients.

Privacy-preserving techniques can be combined with interpretability methods: The approach incorporates "metric privacy techniques that provide formal privacy guarantees while preserving utility," showing that organizations don't have to choose between understanding their models and protecting sensitive information.

This represents a new application of XAI in machine learning: The authors note this is "the first time" feature attribution methods have been applied to federated learning contexts, suggesting XAI's potential extends beyond model transparency to solving concrete technical challenges in collaborative AI systems.