A Framework for Behavior Privacy Preserving in Radio Frequency Signal

Jianwei Liu, Jinsong Han, Lei Yang, Fei Wang, Feng Lin, Kui RenarXiv 2020 · arXiv

TL;DR

RFBP-Net is a deep learning framework designed to "scrub" behavioral information (like gestures or poses) from radio frequency (RF) signals while preserving the features necessary for user authentication. By utilizing a Siamese network architecture, the system reduces activity recognition accuracy by up to 80% while maintaining near-original levels of identity verification, preventing service providers from secretly monitoring user actions.

The Privacy Challenge in RF Sensing

Wireless sensing using WiFi and RFID signals has emerged as a powerful tool for non-intrusive human-computer interaction. It allows systems to identify users and track movements without the need for cameras or wearable devices. However, this "over-sensing" capability introduces a significant privacy loophole.

When a user interacts with an RF-based service for authentication—such as entering a home via gait recognition—the service provider (SP) receives a signal containing both the user's identity and their current behavior. An untrusted SP could exploit this data to monitor daily routines, speculate on personal interests, or capture sensitive motions like typing a password. The core challenge lies in moderating the information revealed by RF signals: providing identity utility while strictly preserving behavioral privacy.

The RFBP-Net Framework

To address this, the researchers developed RFBP-Net, a framework that treats privacy preservation as a feature extraction problem rather than a data-obfuscation task (such as adding noise). The framework follows a structured pipeline to ensure that the final output is "behavior-blind" but "identity-aware."

The end-to-end pipeline of the proposed framework, from source data validation to feature quality assessment.

The process begins with source data validation and training set construction, followed by core processing and a final validation module. This final step uses classic machine learning models (e.g., KNN, SVM) to verify that the resulting features maintain high identity accuracy while minimizing activity recognition.

Methodology: Siamese Network and Feature Decoupling

At the heart of RFBP-Net is a novel Siamese network-based deep model. This architecture is designed to decouple identity features from behavioral features through a dual-loss objective function.

The Siamese network architecture used for feature decoupling, highlighting the dual-loss objective function and behavior-irrelevant output.

The model employs two identical sub-networks with shared weights. During training, it processes pairs of signal samples using two primary loss functions:

  • Contrastive Loss ($LOSS_C$): This forces the network to minimize the distance between features of the same user performing different activities. This effectively teaches the model to treat human movement as a nuisance variable or "noise."
  • Identity Loss ($LOSS_P$): A standard cross-entropy loss ensures the extracted features remain highly discriminative for user classification.

A tunable hyperparameter ($\alpha$) allows system designers to balance this trade-off. A higher $\alpha$ prioritizes hiding behavior (privacy), while a lower $\alpha$ prioritizes authentication accuracy (utility).

Experimental Evaluation

The researchers evaluated RFBP-Net across four distinct data sources to test its robustness and cross-domain applicability:

  • Custom RFID Dataset: Over 4,000 samples from 5 volunteers performing 10 "air-writing" activities.
  • Custom WiFi Dataset: Over 29,000 samples from 10 volunteers performing 10 different gestures.
  • Wiar: An open WiFi dataset with 10 volunteers performing 16 different activities.
  • Widar3.0: A large-scale, open cross-domain gesture recognition dataset.

Key Results and Insights

The framework demonstrated a remarkable ability to decouple identity from action across all datasets.

Activity Suppression vs. Utility Retention

In the RFID system, activity recognition accuracy plummeted from 95% to 25%. In the WiFi system, it dropped by 80%, reducing an attacker's insight to the level of a random guess. Despite this heavy filtering, authentication utility remained robust. The RFID system saw only a 5% drop in identity accuracy, while the WiFi system saw a negligible 1% decrease. On the Wiar dataset, the results were even more pronounced: activity recognition fell to just 4.13% while identity accuracy was maintained at 99.5%.

Technical Insights

  • Privacy through Distillation: Unlike Differential Privacy, which can degrade utility, RFBP-Net uses the Siamese network to "distill" the signal, identifying the core components of a WiFi reflection that remain constant regardless of the user's activity.
  • Activation Function Placement: The authors found that placing standard activation functions (like Sigmoid or Tanh) immediately before the output layer destroyed identity-relevant features. To maintain utility, the final feature extractor output must remain unconstrained.

Conclusion and Future Work

While RFBP-Net provides a strong empirical framework for privacy, the authors acknowledge the need for a more rigorous mathematical derivation for feature derivation in future work. Additionally, while this study focused on retaining authentication while hiding behavior, future iterations could explore preserving other functions, such as indoor tracking or localization, while maintaining the same level of behavioral privacy.

Summary updated

Abstract

Recent years have witnessed the bloom development of the human-centered wireless sensing applications, in which some human information, such as the user's identity and motions, can be retrieved through analyzing the signal distortion caused by the target person. However, the openness of wireless transmission raises increasing concerns on user privacy, since either the human identity or human motion is sensitive in certain scenarios, including personal residence, laboratory, and office. Researchers have reported that commodity WiFi signals can be abused to identify users. To dispel this threat, in this paper we propose a privacy-preserving framework to effectively hide the information of user behaviors in wireless signals while retaining the ability of user authentication. The core of our framework is a novel Siamese network-based deep model, namely RFBP-Net. In this way, wireless sensing reveals user information moderately. We conduct extensive experiments on both the real WiFi and RFID system and open datasets. The experiment results show that RFBP-Net is able to significantly reduce the activity recognition accuracy, i.e., 70% reduction in the RFID system and 80% reduction in the WiFi system, with a slight penalty in the user authentication accuracy, i.e., only 5% and 1% decrease in the RFID and WiFi system, respectively.
BibTeX
@misc{liu2020framework,
  author = {Jianwei Liu and Jinsong Han and Lei Yang and Fei Wang and Feng Lin and Kui Ren},
  title = {{A Framework for Behavior Privacy Preserving in Radio Frequency Signal}},
  year = {2020},
  eprint = {2004.04909},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2004.04909},
}
Figure