CALM: A Ubiquitous Crowdsourced Analytic Learning Mechanism for Continual Service Construction with Data Privacy Preservation
Kejia Fan, Yajiang Huang, Jingyu He, Feijiang Han, Jianheng Tang, Huiping Zhuang, Anfeng Liu, Tian Wang, Mianxiong Dong, Houbing Herbert Song, Yunhuai Liu · Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies · 2025
Leveraging collective capabilities, crowdsourcing has become a popular approach for ubiquitous service construction. The most classical paradigm constructs AI services on the platform using the workers' collected data centrally, leading to widespread data privacy concerns. Federated Learning represents another classical attempt to preserve data privacy through the workers' local training in a distributed and iterative manner. However, due to their inherited reliance on the gradients for back-propagation, they both suffer from the famous problem of "catastrophic forgetting". As incremental demands arise, continual training will significantly compromise previously learned knowledge within the AI model, leading to the degradation of service quality. To overcome these limitations, we propose a ubiquitous Crowdsourced Analytic Learning Mechanism, named CALM. By introducing analytic learning into crowdsourcing with meticulously designed computation and communication protocols, our CALM can effectively achieve data privacy preservation and absolute knowledge memorization. In CALM, the workers are required to locally perform lightweight computations without back-propagation using their collected data. Subsequently, they only need to upload the privacy-preserved results to the platform rather than the raw data. Finally, the platform can recursively aggregate the workers' uploaded results to continually update the AI model for providing ongoing services. Our CALM guarantees that the recursively derived model is fully equivalent to that obtained from centralized training using the complete dataset, fundamentally avoiding catastrophic forgetting. Theoretical analyses are comprehensively presented regarding the validity, security, efficiency, and interpretability of CALM. Extensive experiments using multiple real-world datasets are conducted to validate the state-of-the-art performance of CALM for continual service construction.