Privacy-Preserving Spatial Crowdsourcing in Smart Cities Using Federated and Incremental Learning Approach

Spatial crowdsourcing (SC) systems have emerged as an advanced crowdsourcing paradigm to revolutionise the efficient development of smart city services. SC engages participants and their sensitive data to accomplish spatiotemporal tasks on platforms. However, revealing sensitive data in SC for smart cities exacerbates cybersecurity and privacy concerns, especially Membership Inference Attacks (MIA). To address the problems, this research proposes a Federated Learning (FL) and Incremental Learning (IL) based framework in SC that integrates advanced privacy-preserving techniques. By leveraging FL and adaptive differential privacy, sensitive data remains in decentralised devices while local models are trained without exchanging raw data to a server. We integrate additive secret sharing, a secure multi-party computation technique to protect data during transmission and aggregation. IL enhances the framework using a generative replay approach to ensure continuous adaptation to new data without forgetting existing knowledge to overcome catastrophic forgetting. We broadly evaluate our work against MIA and catastrophic forgetting using Yelp datasets. Compared with other baseline approaches, our experimental results demonstrate that the proposed framework significantly mitigates the risk of MIAs by around 50% and improves forgetting accuracy by up to 13%, thereby providing robust privacy-preserving mechanisms.