Edge-Based Personalized Information Retrieval for Mobile Users Leveraging Federated Learning

Concerns about preserving user privacy in mobile information retrieval have become a serious challenge. This study proposes an innovative framework-combining edge computing with federated learning that performs all personalization and data processing entirely on the user's device. In this architecture, lightweight language models such as DistilBERT and TinyBERT are trained locally, and no raw data ever leaves the device. Only encrypted model weights are exchanged between devices and the central server to update the shared base model. For each user, a dedicated personalization layer (Adapter/LoRA) is instantiated on the same device to precisely address individual needs. Empirical evaluation on a synthetic dataset and the MS MARCO collection demonstrated that our framework achieves a precision of ۰.۹۱۲ and a recall of ۰.۸۸۵, while reducing average response latency to just ۲۱ milliseconds. Moreover, model memory usage stays around ۵۲ MB on average, and privacy preservation remains at ۱۰۰ percent throughout all stages. These results show that our edge-based design not only boosts search speed and quality but also runs reliably and efficiently on mobile and IoT devices which offers a practical solution for applications with strict data-sensitivity requirements.

کلیدواژه ها:

Edge Computing ، Information Retrieval ، Federated Learning ، Personalization ، Privacy-Preserving Search