HyperFed: Free-riding Resistant Federated Learning with Performance-based Reputation Mechanism and Adaptive Aggregation using Hypernetworks

Traditional machine learning solutions rely on Cloud based services, which could potentially lead to major problems including security, privacy data leakage, unacceptable latency, and excessive operating expenses. Federated Learning techniques (FL) were introduced to tackle these challenges by allowing distributed edge nodes/servers to collaboratively train AI models without sharing raw training data. However, some of the nodes may intentionally or unintentionally upload virtual (fake) models to the main server. This behavior is called "Free-riding", and it could potentially have a negative effect on the overall performance of the FL system. In this paper, we propose a new adaptive contribution-based aggregation technique using hypernetworks, namely "HyperFed", and evaluate it on two important aspects: resistance against free-riders' fake contributions, and average convergence speed of global model on local datasets. Our simulation results on Federated EMNIST dataset display promising performance in comparison to FedAvg and AdaFed aggregation techniques.