This blog documents several trending methods related to security and smart sensing.

Disclaimer:
There are terms that I am unfamiliar with since the paper sort of lies out side of my current research experience but I will try to explain as clear with plain words : )

Overview

Wireless sensing researchers aim at describing the status of surrounding objects by the reflection, scattering, and refraction of wireless signals (e.g., Wi-Fi, RFID and LoRa). On the flip side, wireless sensing could also be used for malicious purpose and that is why there are more and more wireless sensing works published in security conferences.

The common workflow consists of 5 steps: Raw-> Pre-processing-> Feature extraction-> Model formulation-> Application Especially for Pre-processing-> Feature extraction, besides traditional ML, some utilize deep learning, which is seen as “feature agonistic”. From what I’ve seen, the DL-based approach helps with some traditional methods cannot achieve but DL is not the cure-all. The lack of interpretability unsettles security researchers. It is dangerous for a security project to not give convincing evidence but a stunning evaluation result. So before employing DL for wireless security, we should figure out which role DL plays in the workflow.

Pre-processing

Model-based denoising

In the simplest form, to eliminate the reflections from static backgrounds, we can measure the interference initially and subtract it for the following collected signals ¹. For the classic multi-path problem, the simplest way to remove the offset is to compute the power of a received packet ². But we lose ToF and AoA in this way. To address this problem, researchers found that the phase offset presents a linear pattern across subcarriers because the initial phase of a receiver is constant ³. As a result, we can use models like linear regression to reduce noises.

Sometimes multi-path can also be an aid. In ⁴, researchers leverages signals received by multiple antennas(MIMO) of a single AP. They used triangles created by reflections (multi-path triangle). If there is only one triangle, we can only get the AoA and relative ToF (relative among edges). However, with multiple antennas, they can do the geometry.

DL-based denoising

This is usually mixed with feature extraction because DL-based denoising recognizes the background by training. However, what DL model does here is dealing with environmental noise.

Many researches leverage adversarial network to eliminate the influences cast by the environment: A min-max game is proposed for transferring the feature in cross-scenarios conditions. EI ⁵ and RF-Sleep ⁶ integrate the adversarial architecture into the feature extractor, which can learn environmental-independent components by specially designed loss functions for the generator and discriminator as the penalty.

Feature extraction

ML-based

Ideally, we can directly obtain the whole picture. For example, Widar3.0 ⁷ designs the environment-independent Body-coordinate velocity Profile (BVP) with CSI measurements by transforming global coordinate to body coordinate. However, wireless measurements are inherently massive but noisy.

The traditional way of feature extraction are usually the results of empirical study, which mostly, serve single purpose. For example, CSI power ² for path length change. The empirical-based methods are highly explainable because every step of lemma inference is provided by concrete experiments. Hodor ⁸ is a typical ML-based wireless security work that covers pre-processing, feature selection and model formulation. It claims that RF signal of car keys have distinct traits to differentiate themselves from even those of the same manufacturer. After having the features, it feeds to a ML model to expect a threshold output.

Another example is geometry-related. With the processed data, triangulations can be utilized for localization. For example ⁹, uses nearby walls of the voice assistant to locate the sound source of the user. This model is acquired by several tines of pre-attempts to let the device know where it is at (find the nearest wall). Then, it uses the direct path and the reflected path to form a triangle. Note that it also works in NLoS.

DL backbone

DL is known for its learning ability in feature extraction. For example, CNN is designed for 2-D, 3-D analysis and RNN, LSTM are for sequential signals. The examples are numerous, but we should note that exploring the design of the neural network cannot only optimize the selection of DL models, but also enhance the interpretability of AI approaches. In particular, I have to mention GAN again. If we put it in the aspect of feature extraction, it can be utilized to learn the hidden connection between the sourcing inputs and the applied outputs. As GAN is becoming prevalent, it is necessary to understand what it does to your model.

A more recent topic is attention module, which lets the neural network concentrate on a subset of its inputs. The basic idea of the attention module is to emulate the working mechanisms of human brain for content awareness and compute a mask for the input feature. For example, ¹⁰ leverages self-attention to allow the model to attend to different joints over time differently, depending on their confidence scores.

Thoughts

• The non-interpretability stalls security researchers to explore the use of DL. But does DL necessarily replace explainable feature engineering? DL gives the whole system a performance gain in average, but also suffers from poor generalizability, which is why GAN is now so popular. However, would GAN be better if it combines with some prior knowledge?

• There are things that DL cannot fix. For example, as DL boosts the accuracy of recognition, the trade-off between sensing granularity and robustness is largely dependent on the sensitivity of signal itself. This actually indicates a new line of research. ** Can different types of signal combine? ** For example, autonomous vehicles use both Lidar and Radar to examine surroundings because light is more sensitive but attenuates more seriously than less sensitive sound.

• Following the autonomous vehicle topic, there are few discussions about its wireless security…