The integration of sensing capabilities into emerging wireless standards, such as 802.11 bf, presents an increasing threat to public privacy. Recent studies have demonstrated that even minor activities, such as finger movements on a keyboard, can be detected by exploiting Wi-Fi Channel State Information (CSI). To mitigate the privacy risks associated with Wi-Fi sensing, prior research has explored methods to disrupt the CSI measurement process by injecting interference into the wireless channel. However, current techniques often inundate the channel with excessive interference, resulting in a significant degradation of the wireless communication link. This paper proposes a spectrally efficient approach by investigating the sensitivity of Wi-Fi sensing to the CSI measurement rate, and designing interference that reduces sensing accuracy while preserving the integrity of the communication link. First, we quantify the accuracy of Wi-Fi-based keystroke recognition in relation to the CSI data rate. We then present theoretical justifications for the use of interference and demonstrate its impact on CSI data rates. Then, a method of adversarial interference is applied, reducing Wi-Fi sensing accuracy by 70 % in a keystroke detection scenario. Finally, a comprehensive trade-off study is conducted to demonstrate how interference can be optimized to protect privacy with savings of up to 21 % in bandwidth, ensuring minimal degradation of service quality.

Abstract.Traffic infrastructure safety is vital to society, and critical traffic corridor breakdowns could halt cities. Particularly, vehicular tunnels are notable due to the constrained space and challenges to emergency response. Existing tunnel monitoring solutions, such as cameras and thermal sensors, are prone to false positives and usually expensive. To address these problems, we investigate the efficacy of Internet-of-Things (IoT) enabled WiFi sensing technology for environmental safety monitoring using versatile, low-cost embedded devices. As a proof-of-concept, we deployed our WiFi sensing devices in an urban underwater tunnel to monitor the tunnel temperature and fire accidents. Specifically, we set up one long-term deployment for temperature monitoring and conducted two fire detection experiments in the tunnel using a small-scaled bonfire and an actual vehicle fire. Our experiments show that the proposed WiFi sensing system could accurately monitor the ambient tunnel temperature to an error of≤0.4∘C during the long-term deployment. In the smallscaled bonfire experiment, we developed a novel empirical model to characterise fire intensity using WiFi signals. This proposed linear regression model has a coefficient of determination of 0.76. Finally, we validate the potential of our wireless environment vision system for detecting fire events from ambient environments using a low-complexity machine learning classifier enabling in situ processing locally on IoT devices.

The rise of Wi-Fi sensing applications leveraging Channel State Information (CSI) from ambient wireless signals has opened up extensive opportunities for human activity and identity recognition. However, this advancement raises serious privacy concerns, as sensitive personal data can be inferred by applying advanced Machine Learning (ML) algorithms to CSI data. In response, researchers have explored adversarial techniques to degrade Wi-Fi sensing accuracy and protect privacy, often by interfering with or corrupting CSI. This paper introduces Wi-Spoof, a novel approach for spoofing CSI to deceive Wi-Fi-based Human Activity Recognition (HAR) systems. Wi-Spoof manipulates Wi-Fi transmission power to inject noise into the CSI and employs a pseudo-Pulse Width Modulation (PWM) scheme to generate controlled, adversarial CSI. Using commercially available hardware, we experimentally demonstrate that Wi-Spoof can achieve targeted misclassification in a state-of-the-art HAR system with a 93% success rate. Our approach is validated on a widely recognised public dataset and further supported by extensive local experiments, underscoring Wi-Spoof’s effectiveness in steering HAR predictions to specified outcomes.

Road tunnel safety is vital to society, and break-downs due to fire accidents could halt cities. Existing tunnel monitoring solutions, such as cameras and thermal sensors, are prone to false positives and usually expensive. To address these problems, we investigate the efficacy of our WiFi sensing technology for environmental safety monitoring in a versatile and low-cost manner using commercial WiFi-enabled IoT devices. As a proof-of-concept, we conducted both a small-scale bonfire experiment and a real vehicle fire detection experiment in a traffic tunnel using our proposed WiFi environment vision system. As a result, we verified that WiFi sensing can be used to monitor fire accidents from the bonfire experiment, and developed a novel empirical model to characterise car fire intensity using WiFi signals with a coefficient of determination of 0.319. Additionally, we also demonstrated the potential for detecting fire outbreaks using a low-complexity machine learning-based classifier, capable of identifying the occurrence of fire accidents on the edge.

With the increasing popularity of WiFi in recent years, WiFi-based wireless sensing technologies have attracted tremendous research. As commercial WiFi networks transition from the 2.4GHz band into the 5GHz band, no prior work has explicitly investigated the impact that this choice of frequency has on sensing outcomes. On both frequency bands, this paper uses the frequency selective behaviour of CSI and a support vector machine classifier to verify the accuracy of human surveillance applications. These experiments demonstrate that 5GHz WiFi offers superior sensing outcomes, with a 6% increase in human occupancy counting accuracy. By restricting the domain of Orthogonal Frequency Division Multiplexing (OFDM) subcarriers, we perform further experiments to conclude that the increase in accuracy is primarily a result of having the higher number of subcarriers or the larger bandwidth for 5GHz WiFi. To corroborate the robustness of our sensing system, we conduct the experiments over multiple trials for two very different environments; one being a controlled elevator setup, while the other being a more realistic and uncontrolled, office space. This novel experimental investigation motivates the need for conducting more comparison studies across the electromagnetic spectrum to identify the application specific best frequency spectrum for the next generation of wireless sensing technologies.

WiFi sensing has gained tremendous traction due to its inherent advantages in terms of privacy and ubiquity. Recent work has shown the ability to sense physical environments, such as counting the number of human occupants. These results have traditionally been achieved using statistical features on WiFi Channel State Information (CSI) amplitude, however more recently there has been interest in exploiting Image based Machine Learning (ML) techniques to achieve better outcomes. In this work, we produce Probability Mass Function (PMF) Images on WiFi CSI, to create spectral maps which clearly distinguish between different human occupancies. We validate our PMF images with common default CNN architectures such as GoogleNet, ResNet and ShuffleNet. By changing the filter size and training parameters, we improve the performance of ShuffleNet from 84% to 98%. Furthermore, we demonstrate how the PMF images can be optimised for sensing outcomes, by controlling the image resolution.

WiFi Sensing has received tremendous attention in Recent Literature, demonstrating the ability to leverage ubiq-uitous commercial WiFi devices to sense Human activities and environmental occupancy. We identify that in all environments fire-safety is vital, and this paper demonstrates the suitability for using WiFi to sense fire. Using commodity Raspberry Pi devices on the 5GHz WiFi band we demonstrate a temporal shift in WiFi Channel State Information (CSI) Amplitude, before, during, and after the ignition of a flame. We further emphasise the presence of fire by observing the spread of CSI Amplitudes, noting that CSI takes much more diverse values in the presence of fire. This result is exacerbated by the frequency selective behaviour of OFDM subcarriers, where some subcarriers displayed larger variation in CSI amplitude due to the fire. The WiFi Fire sensing model was evaluated in an ideal setup with a gas flame to remove material deformation as a variable, and subsequently in a real-world scenario with the ignition of building cladding.

Extensive literature has shown the possibility of using WiFi to sense large scale environmental features such as people, movement, and human gestures. To our best knowledge, there has been no investigation on identifying the microscopic changes in a channel due to atmospheric temperature variations. We identify this as a real world use case, since there are scenarios such as Data Centres where WiFi traffic is omnipresent and temperature monitoring is important. We develop a framework for sensing temperature using WiFi Channel State Information (CSI), proposing that the increased kinetic energy of ambient gas particles will affect the wireless link. To validate this, our paper uses low wavelength 5GHz WiFi CSI from commodity hardware to measure how the channel changes as the ambient temperature is raised. Empirically, we demonstrate that the CSI amplitude value drops at a rate of 13 per degree Celsius rise in the ambient temperature based on the testing platform, and developed regressions models with ± 1°C accuracy in the majority of cases. Moreover, we have shown that WiFi subcarriers exhibit a frequency-selective behaviour in their varying responses to the rise in ambient temperature.

Due to the increased proliferation of WiFi in public and private spaces, there is interest in exploiting WiFi for spatial monitoring. In this paper, we aim to characterize movement or objects in a channel using Channel State Information (CSI). Channel state information represents the degree to which a wireless signal has been attenuated and delayed, and hence we hope to characterize different objects and multipath channel characteristics from CSI. We place different static objects and moving humans in a channel and inspect the CSI for each channel condition. From the variations in CSI Amplitude we can accurately distinguish between a person walking, squatting, or standing still in the channel. To identify static objects, we present a novel approach by inspecting the CSI of different Orthogonal Frequency Division Multiplexing (OFDM) subcarriers. This paper makes a novel contribution, by observing frequency selective behavior of CSI for different channel stimuli. This can be used to improve channel detection accuracy.