The article introduces ProvFL, a mechanism that aids in interpreting and explaining global model predictions in Federated Learning (FL). It’s designed to track the information flow between individual clients in FL and the final global model, focusing on influential and sensitive neurons. The tool helps understand what clients contribute to a global model and who is responsible for a particular prediction. It uses the invertible nature of fusion algorithms to determine each client’s contribution. ProvFL shows an average provenance accuracy of 97% and outperforms other FL fault localization approaches by an average margin of 50%.
Publication date: 21 Dec 2023
Project Page: https://arxiv.org/abs/2312.13632
Paper: https://arxiv.org/pdf/2312.13632
