Passive Activity Classification Using Just WiFi Probe Response Signals

Abstract

Passive WiFi radar shows significant promise for a wide range of applications in both security and healthcare owing to its detection, tracking and recognition capabilities. However, studies examining micro-Doppler classification using passive WiFi radar have relied on manually stimulating WiFi access points to increase the bandwidths and duty-cycles of transmissions; either through file-downloads to generate high data-rate signals, or increasing the repetition frequency of the WiFi beacon signal from its default setting. In real-world scenarios, both these approaches would require user access to the WiFi network or WiFi access point through password authentication, and therefore involve a level of cooperation which cannot always be relied upon e.g. in law-enforcement applications. In this research, we investigate WiFi activity classification using just WiFi probe response signals which can be generated using a low-cost off-the-shelf secondary device (Raspberry Pi) eliminating the requirement to actually connect to the WiFi network. This removes the need to have continuous data traffic in the network or to modify the firmware configuration to manipulate the beacon signal interval, making the technology deployable in all situations. An activity recognition model based on a convolutional neural network resulted in an overall classification accuracy of 75% when trained from scratch using 300 measured WiFi proberesponse samples across 6 classes. This value is then increased to 82%, with significantly less training when adopting a transfer learning approach: initial training using WiFi data traffic signals, followed by fine-tuning using probe response signals.

Transfer Learning for Future Wireless Networks: A Comprehensive Survey

“…In practice, some emergency services, e.g., search-rescue applications, may have urgent cooperation with the wireless network without using the beacon signals. For that, the authors in [157] present a TL-based passive activity recognition using only Wi-Fi probe response signals generated from a local Wi-Fi device, e.g., Raspberry Pi, without connecting to the actual Wi-Fi network. Specifically, the probe request-response mechanism of Wi-Fi can enable the information exchanges between the local Wi-Fi AP and its users without any encryption or modification to the AP.…”

Transfer Learning from Audio Deep Learning Models for Micro-Doppler Activity Recognition

“…The effectiveness of this knowledge transfer is dependent on the extent of similarity between the two tasks. For µ-D signature classification, only transfer learning from CNNs trained on images has been attempted [11,12]. This paper would be the first to conduct transfer learning from a CNN trained on audio for this task.…”

“…The µ-D dataset used in the experiments was obtained from a passive Wi-Fi setup detailed in [12]. The dataset comprises of 1,109 spectrograms across six movement classes: taking a bow, performing a breast-stroke motion, performing a crawl-stroke motion, performing a double punch, sitting, and standing.…”

Self‐supervised multimodal fusion transformer for passive activity recognition

“…Deep neural network architectures that are commonly used for Wi-Fi CSI-based sensing include convolutional neural networks (CNN) where the raw CSI data is transformed into image-like representations such as spectrograms (e.g. [16][17][18]22]) or recurrent neural networks (RNN) which work directly on the raw Wi-Fi data (e.g. [23,24]).…”