PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative AI based Secure Wireless Sensing for ISAC Networks

Jiacheng Wang, Hongyang Du, Yinqiu Liu, Geng Sun, Dusit Niyato, Shiwen Mao, Dong In Kim, Xuemin Shen

arXiv:2408.11398v1eess.SP

TL;DR

CSI-based sensing creates security risks because unauthorized devices can infer user activities from intercepted signals. DFSS uses conditional diffusion models for link selection and safeguarding-signal generation, then modulates the signals onto pilots. In experiments, DFSS reduced unauthorized activity-recognition accuracy by approximately 70%.

  • Problem

    CSI-based sensing can expose users to unauthorized sensing, while existing ISAC research has largely emphasized sensing performance rather than this security issue.

  • Method

    DFSS uses discrete conditional diffusion for graph-based link and node activation, continuous conditional diffusion for safeguarding signals, and pilot modulation to mask activity-related CSI fluctuations.

  • Results

    Approximately 70%: DFSS reduces unauthorized devices’ activity-recognition accuracy.

  • Takeaways & Limitations

    DFSS demonstrates the potential of generative AI to protect users from illegitimate wireless sensing while preserving authorized CSI extraction.

Abstract

from arXiv · show

Integrated sensing and communications (ISAC) is expected to be a key technology for 6G, and channel state information (CSI) based sensing is a key component of ISAC. However, current research on ISAC focuses mainly on improving sensing performance, overlooking security issues, particularly the unauthorized sensing of users. In this paper, we propose a secure sensing system (DFSS) based on two distinct diffusion models. Specifically, we first propose a discrete conditional diffusion model to generate graphs with nodes and edges, guiding the ISAC system to appropriately activate wireless links and nodes, which ensures the sensing performance while minimizing the operation cost. Using the activated links and nodes, DFSS then employs the continuous conditional diffusion model to generate safeguarding signals, which are next modulated onto the pilot at the transmitter to mask fluctuations caused by user activities. As such, only ISAC devices authorized with the safeguarding signals can extract the true CSI for sensing, while unauthorized devices are unable to achieve the same sensing. Experiment results demonstrate that DFSS can reduce the activity recognition accuracy of the unauthorized devices by approximately 70%, effectively shield the user from the unauthorized surveillance.

I. INTRODUCTION

CSI-based sensing supports simultaneous wireless communication and sensing but exposes users to unauthorized surveillance and spoofing attacks. DFSS addresses these risks with diffusion-based graph and safeguarding-signal generation, reducing unauthorized activity-recognition accuracy by about 70%.

  • ISAC integrates wireless communication and environmental sensing, while CSI-based sensing can infer users’ locations, activities, breathing, and heart rates.
  • Unauthorized devices can intercept standardized training symbols, estimate CSI, and infer individuals’ activities, creating substantial privacy risks.
  • CSI-based sensing also faces rogue-AP attacks that alter signal amplitude and phase to disrupt legitimate sensing and communications.
  • DFSS uses a discrete conditional diffusion model to generate graphs that activate wireless links appropriately while maintaining sensing performance and reducing operation cost.
  • DFSS generates diverse, random safeguarding signals with a continuous diffusion model and modulates them onto pilot symbols to mask user-activity fluctuations.
  • Approximately 70%: DFSS reduces unauthorized devices’ activity-recognition accuracy in software-defined-radio experiments.

B. Security Issues in CSI Based Sensing

CSI-based sensing is vulnerable to unauthorized inference and attacks on legitimate sensing. DFSS masks activity-related CSI fluctuations by generating and modulating safeguarding signals, while authorized receivers retain access to the true CSI.

  • Security Issues in CSI Based Sensing: Unauthorized devices can eavesdrop on CSI to infer user information, while rogue systems can inject spoofing signals that cause incorrect sensing decisions.
  • Security Issues in CSI Based Sensing: DFSS generates a safeguarding signal and modulates it onto the pilot signal to mask fluctuations caused by user activity.
  • Security Issues in CSI Based Sensing: DFSS uses three modules: graph generation, safeguarding-signal generation, and modulation onto the transmitted pilot.
  • Signal Model: In the CSI model, pilots are transmitted across OFDM subcarriers, and received training symbols relate to pilots and channel gains through Y = X×H.
  • Signal Model: CSI power contains static and dynamic path components, with activity-related fluctuation characteristics primarily determined by dynamic-path terms.
  • Security Issues in CSI Based Sensing: Authorized receivers use shared safeguarding signals to estimate actual CSI, whereas unauthorized receivers lack the signal needed to extract it correctly.

C. Graph Generation for Link Selection

DFSS represents transmitters and receivers as graph nodes and sensing links as edges, conditioning graph generation on device distribution and user location. The generated graph guides economical activation of sensing links and devices.

  • C. Graph Generation for Link Selection: DFSS treats signal transmitters and receivers as nodes, sensing links as edges, and device distribution and user location as graph-generation conditions.
  • C. Graph Generation for Link Selection: The generated graph guides activation of appropriate wireless links and sensing devices to support effective and economical system operation.

1) Problem Formulation:

CSI-based sensing uses dynamic-path information to characterize user activity, but selecting sensing links and devices must balance sensing performance against activation cost.

  • The dynamic path contains information about user activity, making SSNR a measure of sensing performance.SSNR is defined as the ratio of dynamic-path power to noise plus static-path power; higher SSNR indicates better sensing performance.
  • SSNR depends on the distances between the user and the signal transmitter and receiver.The formulation also includes transmission power, antenna gain, reflection area, and transceiver-pair constants.
  • The optimization selects sensing links, transmitters, and receivers while maintaining sensing performance with fewer activated wireless links.The objective incorporates the aggregate SSNR together with link and node activation costs through weighting factors.

2) Graph generation:

D-CDM generates an activation graph conditioned on device and user locations, then learns its denoising policy to optimize sensing utility while producing deployable link and node selections.

  • Graph representation: D-CDM represents the activation graph as node and edge types encoded in matrices K and E, producing G0 = (K0, E0).Its discrete diffusion process operates over nodes and edges rather than image pixels.
  • Conditional denoising network: The denoising network uses device distribution and user location as conditions and a noisy graph as input to generate a clear graph.The graph transformer architecture processes node and edge embeddings with attention-based graph-transfer layers.
  • Policy optimization: D-CDM models graph denoising as a Markov decision process, where states, actions, transitions, and rewards define the graph-generation trajectory.The trajectory from GT to G0 is treated as a sequence of state-action pairs, with cumulative reward equal to the final graph reward.
  • Policy optimization: Monte Carlo estimation approximates the policy gradient, while grouping trajectories by equivalent final graphs simplifies optimization.The implementation samples 256 trajectories and timesteps randomly from 1 to 30.
  • Graph activation: After training, the denoising network generates graphs from new conditions to guide activation of sensing links and nodes.Network parameters are updated using the estimated policy gradient and a learning rate η.

D. Safeguarding Signal Generation

After link and node activation, the continuous diffusion model is trained on a dataset conditioned on user locations to generate diverse safeguarding signals.

  • The continuous diffusion model generates safeguarding signals conditioned on user locations and the selected graph-generation output.Unlike the graph stage, this model is trained with a dataset rather than a reward function.
  • Generated safeguarding signals resemble but differ from training samples, providing diversity and randomness.This distinction is intended to prevent unauthorized receivers from replicating the signal-generation process.

1) Continuous Diffusion Model:

The continuous conditional diffusion model perturbs safeguarding samples with scheduled noise and learns a reverse Markov chain to regenerate signals from noise under specified conditions.

  • Forward diffusion: The forward process adds prescribed noise over T′ steps so the final state approximates N(0, I).The process uses noise scales β1 through βT′ and identity covariance I.
  • Reverse diffusion: The reverse process is a condition-parameterized Markov chain that denoises Gaussian noise into safeguarding signals.The denoising network is trained with a re-weighted evidence lower bound to match the training-sample distribution.
  • Generated-signal properties: The generated signals are similar to but distinct from training signals, helping prevent unauthorized receivers from replicating the process.The resulting variation provides diversity and randomness in the safeguarding signals.

2) Safeguarding Signal Design and Dataset Construction:

The paper constructs composite safeguarding signals from four activity-specific components and trains a continuous diffusion model conditioned on system and user locations. The activity ranges are selected to cover most indoor motion conditions and support masking of activity-induced CSI fluctuations.

  • Safeguarding Signal Design and Dataset Construction: The continuous diffusion model requires a dataset of composite safeguarding signals for training.The paper notes that an ISAC system cannot predict user actions, motivating a composite signal rather than activity-specific generation.
  • Safeguarding Signal Design and Dataset Construction: For sitting down, vl is approximately 0.2–1 m/s, whereas falling spans about 0.8–1.5 m/s and exceeds the other three activities.These statistics were measured in a typical indoor scenario for four activities.
  • Safeguarding Signal Design and Dataset Construction: Four activity-specific safeguarding components correspond to sitting, walking, running, and falling.These components are combined into the composite signal s(fn, t).
  • Safeguarding Signal Design and Dataset Construction: Activity-induced length-change rates span approximately 0 m to 5.5 m across the selected activities, covering almost all indoor scenarios.The range motivates using the four activities to cancel fluctuations from most indoor activities and introduce additional interference.
  • Safeguarding Signal Design and Dataset Construction: The dataset aggregates computed s values from CSI collected for four activities under varied conditions, using those conditions as labels.Figure 5 depicts this construction process.

E. Safeguarding Signal Modulation

DFSS modulates generated safeguarding signals onto transmitted pilots to mask activity-related CSI fluctuations. Authorized receivers regenerate the same signal and recover usable CSI, whereas unauthorized receivers estimate channels from the unmodified pilot.

  • E. Safeguarding Signal Modulation: The generated safeguarding signal is represented as a length-W sequence and applied to the original pilot at each transmission time.The original pilot ±1 + 0i becomes ±1 × s(fn, w) + 0i at time w.
  • E. Safeguarding Signal Modulation: Safeguarding signals assigned to different pilot frequencies are identical at a given time w.This makes the pilot’s time-domain amplitude follow the generated safeguarding-signal trend.
  • E. Safeguarding Signal Modulation: Authorized receivers use the same model and random seed to regenerate s and recover X′ for effective CSI estimation.The recovered CSI is used for communication and sensing.
  • E. Safeguarding Signal Modulation: Unauthorized receivers estimate channels using the original pilot, producing Ĥ′ = X−1Y′ and preventing illegitimate sensing.The modulation process is illustrated in Figure 6.

IV. IMPLEMENTATION AND EVALUATION

The evaluation uses servers, USRP N321 devices, GNU Radio, and optical-fiber or 10-gigabit Ethernet connections. Experiments assess graph generation, safeguarding-signal generation and modulation, and protection against unauthorized activity monitoring.

  • IMPLEMENTATION AND EVALUATION: The experiments evaluate graph generation, safeguarding-signal generation and modulation, and protection against unauthorized monitoring.Human activity recognition is used as the case study for monitoring protection.
  • IMPLEMENTATION AND EVALUATION: Training and inference run on a server, while generated signals are transmitted through a second server connected to USRP devices.The setup uses Ubuntu 20.04, an AMD Ryzen Threadripper PRO 3975WX processor, an NVIDIA RTX A5000 GPU, and GNU Radio 3.
  • IMPLEMENTATION AND EVALUATION: D-CDM takes device and user locations as input and outputs a graph specifying wireless links, transmitters, and receivers to activate.The generated graph is evaluated by a reward function, whose feedback optimizes network parameters.
  • IMPLEMENTATION AND EVALUATION: The continuous diffusion model is trained using a dataset containing composite safeguarding signals.During testing, activated links, nodes, and user location condition the continuous model.

3) Evaluation Metrics:

The evaluation examines graph rewards, denoising behavior, conditional safeguarding-signal generation, and signal quality against GAN and VAE baselines. DFSS produces condition-dependent, diverse signals and achieves stronger SSIM and lower FID than both baselines.

  • Graph Generation: D-CDM reaches an average reward of about 56.16 after roughly 100 epochs, exceeding random and greedy baselines.The node-based and link-based random methods reach approximately 8.84 and 2.77, while greedy methods reach 27.58 and 25.55.
  • Graph Generation: After 30 denoising steps, D-CDM produces an optimal graph whose activated links form a Fresnel zone around the user.Fewer than 12 steps leave noisy nodes and edges, while 18 steps noticeably reduce ineffective elements.
  • Safeguarding Signal Generation: After 500 denoising steps, generated safeguarding signals resemble training trends while remaining distinct from the training signals.Signals remain noisy below 460 steps and become clearer and more structured as denoising proceeds.
  • Safeguarding Signal Generation: Different generation conditions produce unique signal strengths and trends; closer Tx–Rx pairs yield weaker signals than farther pairs.The paper attributes this difference to the amplitude and fluctuation characteristics captured by the receiver.
  • Safeguarding Signal Generation: The generated signals’ diversity and randomness complicate replication by unauthorized devices, supporting user protection.The paper links this property to signals varying with input conditions.

3) Protection Performance in Activity Recognition:

DFSS substantially degrades unauthorized activity-recognition performance across activities and systems, while maintaining consistent protection across communication speeds. Its diverse safeguarding signals produce variable CSI features and unpredictable recognition outcomes.

  • Overall protection: Average ADRs with DFSS reach 0.82, 0.79, 0.74, and 0.70 for AF-ACT, ABLSTM, PhaseAnti, and CeHAR, respectively.These values are reported across five activities and compared with Secur-Fi.
  • Activity and system effects: DFSS remains applicable across activities, with AF-ACT ADRs of 0.81, 0.83, 0.82, 0.79, and 0.72 for walking, waving, sitting, squatting, and falling.Its impact varies by recognition system, affecting AF-ACT and ABLSTM more than PhaseAnti and CeHAR because the former rely solely on CSI amplitude.
  • Activity and system effects: With DFSS, ABLSTM and CeHAR recognition accuracies across five activities range from 16% to 26% and from 24% to 34%, respectively.The confusion matrices also indicate that DFSS does not cause systems to classify different activities into the same one.
  • Communication-speed robustness: Across transmission rates from 100 to 600, AF-ACT ADRs remain between 0.78 and 0.81, while CeHAR ADRs remain between 0.69 and 0.72.The paper attributes this consistency to resampling of the generated safeguarding signals.
  • Overall protection: About 70%: DFSS reduces unauthorized activity-recognition accuracy, demonstrating protection against illegitimate sensing.The evaluation uses software-defined radio devices and activity recognition as the example task.
Loading 2408.11398v1…