WAFFLe: Weight Anonymized Factorization for Federated Learning

In domains where data are sensitive or private, there is great value in methods that can learn a distributed manner without the ever leaving local devices. light of this need, federated learning has emerged as popular training paradigm. However, many approaches trade transmitting for communicating updated weight parameters each device. Therefore, successful breach would have otherwise directly compromised instead grants whitebox access to model, which opens door number attacks, including exposing very seeks protect. Additionally, scenarios, individual client devices commonly exhibit high statistical heterogeneity. Many common single global model; while may do well on average, performance degrades when i.i.d. assumption violated, underfitting individuals further from mean and raising questions fairness. To address these issues, we propose Weight Anonymized Factorization Federated Learning (WAFFLe), an approach combines Indian Buffet Process with shared dictionary factors neural networks. Experiments MNIST, FashionMNIST, CIFAR-10 demonstrate WAFFLe’s significant improvement test fairness simultaneously providing extra layer security.