VPrWS: Vision transformer-based prototypical network with cross-inductive bias distillation for cross-domain wireless human sensing

Abstract

Wireless-based human sensing (WHS) is a powerful technique for device-free activity recognition. It analyzes channel state information (CSI) extracted from wireless signals. However, domain shift remains a major challenge, as performance degrades significantly when models are applied across users, environments, or locations. Few-shot learning (FSL) offers a practical solution by enabling rapid adaptation using limited labeled samples. Among FSL approaches, Prototypical Networks (PNs) are favored for their simplicity and effectiveness in low-data regimes. However, due to the heterogeneous nature of CSI data, constructing a robust and semantically meaningful prototype metric space remains challenging.

To address this issue, we propose VPrWS, a system that integrates a prototypical network (PN) with a vision transformer (ViT) and employs cross-inductive-bias knowledge distillation to enrich the prototype space. Specifically, we develop a knowledge distillation framework for PNs that induces complementary inductive biases from bidirectional long short-term memory (BiLSTM)- and convolutional neural network (CNN)-based PNs to a ViT-based PN. This approach yields a more robust and semantically structured metric space, especially in the context of CSI data, thereby improving cross-domain WHS performance without materially increasing inference time. Evaluations on three public CSI datasets under both in-domain and cross-domain settings demonstrate that VPrWS consistently outperforms existing representative baselines.