CCS: Continuous Learning for Customized Incremental Wireless Sensing Services

TL;DR

CCS tackles the challenge of providing continuous customized wireless sensing services while protecting privacy. It introduces exemplar-based rehearsal, knowledge distillation from a frozen older model, and weight aligning to prevent catastrophic forgetting during on-device updates. On the XRF55 dataset across Wi-Fi, mmWave radar, and RFID, CCS outperforms baselines such as iCaRL, UCIR, BiC, and OneFi, achieving higher ACCN values and sustained accuracy as new classes are added. This work informs future business and system design for privacy-preserving, scalable wireless sensing services and points to practical integration with federated learning and on-device computation.

Abstract

Wireless sensing has made significant progress in tasks ranging from action recognition, vital sign estimation, pose estimation, etc. After over a decade of work, wireless sensing currently stands at the tipping point transitioning from proof-of-concept systems to the large-scale deployment. We envision a future service scenario where wireless sensing service providers distribute sensing models to users. During usage, users might request new sensing capabilities. For example, if someone is away from home on a business trip or vacation for an extended period, they may want a new sensing capability that can detect falls in elderly parents or grandparents and promptly alert them. In this paper, we propose CCS (continuous customized service), enabling model updates on users' local computing resources without data transmission to the service providers. To address the issue of catastrophic forgetting in model updates where updating model parameters to implement new capabilities leads to the loss of existing capabilities we design knowledge distillation and weight alignment modules. These modules enable the sensing model to acquire new capabilities while retaining the existing ones. We conducted extensive experiments on the large-scale XRF55 dataset across Wi-Fi, millimeter-wave radar, and RFID modalities to simulate scenarios where four users sequentially introduced new customized demands. The results affirm that CCS excels in continuous model services across all the above wireless modalities, significantly outperforming existing approaches like OneFi.

Figures (8)

• Figure 1: We present CCS. With CCS, the service provider can not only provide base wireless sensing service for users but also provide incremental model service to meet users' continuous customized sensing demand under the premise that users' wireless sensing data is never transmitted to the service provider to ensure data privacy protection.
• Figure 2: We incorporate knowledge distillation and weight aligning throughout the training process of each incremental stage. During each incremental stage, exemplars are extracted from the training data of the previous stage using the Herding method Herding. The network learns to retain knowledge of old tasks by employing knowledge distillation from a frozen model of the previous stage, utilizing MSE loss for distillation. Additionally, weight aligning addresses the issue of imbalance in the weight distribution between new and old services.
• Figure 3: ACCN of CCS services for user1 with continuous stages listed in Table. \ref{['tab:user_sequences']}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL icarl, UCIR ucir, BiC bic and OneFi onefi, demonstrating a consistent rise in model value.
• Figure 4: ACCN of CCS services for user2 with the continuous demands listed in Table. \ref{['tab:user_sequences']}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL icarl, UCIR ucir, BiC bic and OneFi onefi, demonstrating a consistent rise in model value.
• Figure 5: ACCN of CCS services for user3 with the continuous demands listed in Table. \ref{['tab:user_sequences']}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL icarl, UCIR ucir, BiC bic and OneFi onefi, demonstrating a consistent rise in model value.