On Device Incremental Learning Algorithms for Real Time Personalization Without Data Storage

Abstract

The fast-paced development of edge computing, intelligent mobile devices, wearable computing, and IoT technologies necessitates the deployment of secure, privacy-aware, and real-time personalization mechanisms. The purpose of this research is to design an effective on-device incremental learning algorithm capable of adapting itself to the changing user behavior without utilizing persistent data storage or a cloud server. Real-time adaptation of the machine learning models with the lowest computational complexities, minimal communication overheads, low memory utilization, and no privacy implications can be achieved using the proposed algorithm. The proposed architecture comprises real-time data acquisition, preprocessing, feature extraction, incremental learning process, optimization process, transient memory-based privacy mechanism, and inference process. Streaming data acquired from the edge devices are locally preprocessed and disposed of in real-time without persistent storage for future analysis. The model was created using the deep learning modules of Python programming language and tested through simulations based on mobile/IoT-based behavioral datasets in edge computing environments. Various performance measures including accuracy, precision, recall, F1-score, root mean square error, latency, memory consumption, and energy consumption were used to evaluate the model's performance. The on-device incremental learning model proposed achieved an accuracy of 96.38%, precision of 95.74%, recall of 95.18%, and F1-score of 95.46%, thereby outperforming traditional static and cloud-based personalization models. In terms of latency, memory utilization, and energy consumption, the model improved from 185 ms, 512 MB, and 6.4 W respectively to 42 ms, 148 MB, and 2.1 W respectively. This study proves that the lightweight on-device incremental learning approach, together with transient memory optimization, offers a viable and effective solution that ensures security, scalability, energy efficiency, and privacy for real-time intelligent personalization applications in edge-driven environments.