Differential Privacy Protection Based on Federated Learning in Mobile Crowdsensing

Mobile Crowdsensing (MCS), as a novel data acquisition paradigm in the Internet of Things (IoT), incentivizes a large number of participants to collaboratively sense data for providing real-time services and accomplishing complex sensing tasks to benefit society. However, a major challenge hindering the further development of MCS is the risk of privacy leakage of participant data. In this paper, an effective integration of Federated Learning (FL) with MCS is proposed. The classic Federated Averaging (FedAvg) algorithm is enhanced, and differential privacy (DP) is introduced to locally preserve the privacy of sensitive user data, referred to as DP-FAG. In DP-FAG, noise is applied to sensitive participant data prior to global model aggregation. Moreover, all sensitive training data is securely stored on participant devices, effectively addressing the trusted third-party issue and ensuring data confidentiality. We conduct extensive experimental analysis on image classification tasks to validate the soundness and effectiveness of our proposed methodology.