Abstract
Radio Frequency (RF) sensing technologies have experienced significant growth due to the widespread adoption of RF devices and the Internet of Things (IoT). These technologies enable numerous applications across healthcare, smart homes, industrial automation, and human-computer interaction. However, the non-intrusive and ubiquitous nature of RF sensing—combined with its environmental sensitivity and data dependency—makes these systems inherently vulnerable not only as attack targets but also as powerful attack vectors. This survey presents a comprehensive analysis of RF sensing security, covering system-level vulnerabilities such as signal spoofing, adversarial perturbations, and model poisoning, alongside the misuse of sensing capabilities for cross-boundary surveillance and side-channel inference. We propose unified threat models to structure these attack vectors and conduct task-specific vulnerability assessments across key RF sensing applications. Finally, we review defense strategies across physical, signal, and model layers to provide a structured view of protection mechanisms.
Introduction
RF sensing is a double-edged sword. While it enables revolutionary touchless applications, it is inherently vulnerable to both digital and physical attacks. As RF sensing moves from specialized radar to ubiquitous Wi-Fi and IoT devices, the security landscape has become fragmented. Most existing research focuses on specific modalities (like Wi-Fi) or specific tasks (like activity recognition), leaving a gap in our understanding of the broader threat surface. This paper addresses the lack of a unified framework for RF security, highlighting a fundamental tension: the very properties that make RF sensing useful—its ability to penetrate obstacles and operate without cameras—also make it a perfect vector for "invisible" privacy invasions.
A Unified Framework for RF Security
To structure the complex landscape of RF threats, this survey introduces two novel, unified threat models that categorize how signals can be "weaponized":
- Intrinsic Vulnerability Exploitation Model (IVEM): This model focuses on attacks directed against the sensing system. Examples include adversarial perturbations designed to fool a gesture recognizer or model poisoning intended to degrade system performance over time.
- Sensing-based Threat Invasion Model (STIM): This model focuses on the malicious use of RF sensing. Here, the technology itself is the weapon, used for unauthorized through-wall surveillance, keystroke inference, or acoustic eavesdropping.
The RF Sensing Pipeline and Vulnerability Layers
The survey conducts a multi-dimensional analysis across three critical layers: Physical (hardware and environment), Signal (modulations and Channel State Information), and Model (deep learning pipelines). By examining the standard RF sensing workflow, we can identify specific points of failure and attack surfaces.
As shown in the workflow, vulnerabilities can be introduced at any stage:
- Data Collection: Attackers can manipulate the physical environment using metasurfaces or "ghost" targets to deceive sensors.
- Data Preprocessing: Noise elimination stages can be bypassed or exploited to hide malicious signal components.
- Feature Extraction & Prediction: Adversarial machine learning techniques can target both hand-crafted and learned features (e.g., CNNs, RNNs) to force incorrect classifications or regressions.
Key Insights: The "Double-Agent" Nature of RF
RF sensing isn't just a target for hackers; it is a powerful attack vector. An attacker can deploy a hidden RF sensor behind a wall to "see" into a private home, effectively bypassing traditional camera-based security. Furthermore, the rise of Integrated Sensing and Communication (ISAC) in 6G will turn every base station into a potential sensor, massively expanding the attack surface.
Other emerging threats include:
- Cross-Modal Interference: Attacks like "IS-WARS" use ambient signals from one modality (e.g., ZigBee) to subtly disrupt sensing in another (e.g., Wi-Fi) without affecting data communication.
- Semantic Privacy Breaches: With the integration of Large Language Models (LLMs), RF sensing is moving from recognizing simple actions (walking) to inferring intent (waiting at an ATM to commit a crime), adding a new layer of semantic risk.
Attack Efficacy and Defense Taxonomy
Quantitative findings across the field demonstrate the severity of these threats. Adversarial attacks (such as FGSM or PGD) have been shown to reduce the accuracy of Doppler-based activity recognition systems by up to 85%. In terms of eavesdropping, RF-based acoustic sensing can now reconstruct speech with enough clarity to pass through-wall word detection tests, achieving high accuracy for sensitive data like numeric passwords.
To counter these threats, the survey categorizes defenses into a three-tier framework:
- Physical Layer: Implementing sensor enhancement and geofencing.
- Signal Layer: Utilizing frequency hopping and beamforming to increase the "cost" and complexity of an attack.
- Model Layer: Employing adversarial training and anomaly detection to harden the AI backend against perturbations.
Future Challenges and Conclusion
Despite progress, several "blind spots" remain. There is a critical lack of industry-wide security standards for RF sensing. Most current models are highly environment-dependent, a weakness that attackers can exploit. As we move toward 6G and AI-driven automated attacks, the need for robust, generalizable, and standardized security protocols becomes paramount. This survey serves as a foundation for researchers to bridge the gap between theoretical models and real-world implications in the evolving landscape of RF sensing.
Summary updated