AIFL: Ensuring Unlinkable Anonymity and Robust Incentive in Cross-Device Federated Learning

While cross-device federated learning offers a privacy-preserving data processing approach for Internet of Things (IoT) devices, it introduces fresh privacy risks and elevated computational expenses. Current solutions prioritize data privacy, often overlooking identity privacy vulnerabilities that may lead to personal information disclosure. In addition, achieving personalized client incentives while protecting identity privacy is a serious challenge. In this paper, we propose AIFL, an unlinkable anonymity and robust incentive mechanism in cross-device federated learning. We design a threshold-based accountable ring signature (TARS) protocol that achieves unlinkable anonymity between identity and model updates and ensures accuracy and cost-effectiveness. AIFL identifies dishonest clients by incorporating a detection mechanism that leverages entropy weighting and cosine similarity for malicious update identification. Moreover, our robust incentive mechanism, grounded in the Stackelberg game, adopts differential pricing strategies that take into account clients anonymity budgets, ensuring equitable rewards. Performance evaluation reveals that AIFL achieves superior accuracy and robustness compared to Krum, even under 50% byzantine attackers. Specifically, AIFL outperforms Krum by 0.32%, 29.1%, 79.17%, and 11.86% on MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets in Non-IID, respectively. Additionally, AIFL effectively imposes costly penalties on dishonest clients and enhances the overall utility for honest clients. Consequently, the proposed AIFL holds significant implications for promoting secure and trustworthy machine learning in the IoT domain.