Decentralized Federated Learning Under Free-riders: Credibility Analysis

With aggregator-free control, blockchain-assisted federated learning (BFL) is acknowledged as an effective decentralized paradigm to address the pitfalls of the conventional centralized FL framework, such as single server failures, client dropouts, etc. However, the integration of FL and blockchain is inherently vulnerable to free-rider attack, i.e., attackers disguise themselves to participate in FL without actual contribution to the model training. To well exploit BFL in real applications, it is necessary to quantitatively analyze the impact of such attack on the credibility of BFL. In this paper, we first carefully examine the fundamental characteristics and limitations of FL and blockchain within the underlying physical network infrastructure. In BFL, the delay incurred in model training, blockchain service and model transmission between clients significantly affects the free-riding success probability. Then, we analyze the delay experienced by transmitting and validating a model in BFL with and without free-riders based on queuing theory. Then, we theoretically derive the free-riding success probability with free-riders. Finally, theoretical analysis and simulation experiments show that the system suffers considerable performance degradation when subjected to free-rider attack, and the dominant factors affecting BFL are identified.