Authors:
Zijian Zhao、Tingwei Chen、Zhijie Cai、Hang Li、Xiaoyang Li、Qimei Chen、Guangxu Zhu
Paper:
https://arxiv.org/abs/2408.10919
Introduction
In recent years, Wi-Fi sensing has emerged as a promising technology due to its advantages such as privacy protection, low cost, and penetration ability. This technology has been applied in various domains including gesture recognition, people identification, and fall detection. However, Wi-Fi sensing models often face challenges related to domain shift, where the model’s performance deteriorates in environments different from the training data. This issue is exacerbated by the limited availability of diverse Wi-Fi sensing datasets, leading to models that overfit to the training set and fail in new scenarios.
To address these challenges, the authors propose CrossFi, a cross-domain Wi-Fi sensing framework based on a Siamese network. CrossFi is designed to perform well in both in-domain and cross-domain scenarios, including few-shot and zero-shot learning, and even in scenarios where the testing set contains new categories not present in the training set. The core component of CrossFi is CSi-Net, a sample-similarity calculation network that uses an attention mechanism to capture similarity information. Additionally, CrossFi includes a Weight-Net that generates templates for each class, enabling the framework to adapt to different scenarios.
Related Work
Wi-Fi Sensing and Domain Shift
Wi-Fi sensing leverages variations in signal strength and multipath propagation to extract valuable information such as Channel State Information (CSI) and Received Signal Strength Indicator (RSSI). This technology has been applied in various fields, including fall detection, gesture recognition, and people identification. However, existing Wi-Fi sensing models often struggle with domain shift, where a model trained in one environment performs poorly in a different environment.
Model-Based and Data-Driven Methods
Wi-Fi sensing methods can be categorized into model-based and data-driven approaches. Model-based methods require significant expertise and are often not easily transferable to other tasks. Data-driven methods, particularly those based on deep learning, can learn directly from data without explicit assumptions about the underlying model. However, these methods face challenges in cross-domain scenarios due to the influence of the environment on Wi-Fi signals.
Siamese Networks
Siamese networks have been proven effective in cross-domain tasks. Unlike traditional neural networks that directly output classification results, Siamese networks calculate the similarity between embeddings of two samples. This approach allows the model to capture the similarity relationship between samples from different domains, making it suitable for one-shot and few-shot learning scenarios.
Research Methodology
Overview of CrossFi
CrossFi consists of two main components: CSi-Net and Weight-Net. CSi-Net is an improved Siamese network that uses an attention mechanism to calculate similarity between samples. Weight-Net generates templates for each class based on the relationship between different samples. The workflow of CrossFi includes data collection, data pre-processing, training, and inference phases.
CSi-Net: Similarity Calculator
CSi-Net enhances the traditional Siamese network by incorporating a multi-attention layer to assess similarity between two inputs. This modification allows the model to capture more valuable information compared to traditional methods that rely on distance or cosine similarity.
Weight-Net: Adaptive Template Generator
Weight-Net generates templates for each class by evaluating the quality of samples using the similarity score matrix output by CSi-Net. This approach ensures high-quality templates, improving the model’s performance in various scenarios.
Experimental Design
Data Collection and Pre-processing
Data is collected from different domains and split into training, support, and testing sets. The pre-processing phase involves computing the amplitude and phase of the CSI, calculating the cosine value of the phase, and interpolating missing CSI positions to ensure consistent data dimensions.
Training Phase
The training phase includes two alternate steps: comparative learning and template learning. In comparative learning, CSi-Net is trained to evaluate the similarity between samples. In template learning, Weight-Net generates templates for each class, and CSi-Net evaluates the similarity between samples and templates.
Inference Phase
In the inference phase, the trained CSi-Net and generated templates are used to calculate the similarity score matrix between testing samples and templates. The classification result is determined based on the highest similarity score.
Results and Analysis
In-Domain Experiment
In the in-domain scenario, CrossFi outperforms previous Wi-Fi sensing models in both gesture recognition and people identification tasks, achieving accuracies exceeding 98%.
Few-Shot Experiment
In the few-shot experiment, CrossFi demonstrates superior performance in cross-domain and new-class scenarios. The model achieves high accuracy with only a few samples per class, comparable to the in-domain scenario.
Zero-Shot Experiment
In the zero-shot experiment, CrossFi significantly outperforms traditional zero-shot learning methods, demonstrating its effectiveness in scenarios where no labeled data from the target domain is available.
Ablation Study
A series of ablation experiments illustrate the efficiency of each module in CrossFi. The results show that the multi-attention module and the adaptive template generation method significantly improve the model’s performance.
Overall Conclusion
CrossFi is a robust framework for cross-domain Wi-Fi sensing tasks, leveraging a Siamese network with an attention-based similarity calculation and an adaptive template generation method. The experimental results demonstrate that CrossFi achieves state-of-the-art performance across various scenarios, including in-domain, few-shot, zero-shot, and new-class scenarios. Future work will explore the application of CrossFi in other fields and further improve the model’s stability in cross-category scenarios.
