Robust Localization of Key Fob Using Channel Impulse Response of Ultra Wide Band Sensors for Keyless Entry Systems

TL;DR

The paper addresses secure localization of a vehicle key fob using UWB CIR features for keyless entry, examining baseline pre-computed features and robustness to adversarial perturbations without adversarial training. It introduces a multi-head self-supervised RBF architecture with a noise-regularized encoder-decoder to learn robust, privacy-preserving embeddings, achieving strong performance under FGSM, BIM, and PGD attacks. The approach yields 6% higher accuracy on clean data and 37–38% improvements under adversarial attacks, with up to 67% robustness gains at certain perturbation levels, and a compact ~82 kB model suitable for microcontroller deployment. This work offers a practical, edge-friendly solution for secure UWB-based keyless entry localization that remains effective under adversarial conditions.

Abstract

Using neural networks for localization of key fob within and surrounding a car as a security feature for keyless entry is fast emerging. In this paper we study: 1) the performance of pre-computed features of neural networks based UWB (ultra wide band) localization classification forming the baseline of our experiments. 2) Investigate the inherent robustness of various neural networks; therefore, we include the study of robustness of the adversarial examples without any adversarial training in this work. 3) Propose a multi-head self-supervised neural network architecture which outperforms the baseline neural networks without any adversarial training. The model's performance improved by 67% at certain ranges of adversarial magnitude for fast gradient sign method and 37% each for basic iterative method and projected gradient descent method.

Figures (4)

• Figure 1: Adversarial robustness of our RBF-L4 neural network with multi-head model. X-axis indicates the magnitude of the adversarial noise ($\epsilon$), y-axis indicates the accuracy of the model. Ideally, higher the accuracy even at higher values of $\epsilon$ is desired. The values indicate improved performance of our model over standard architectures.
• Figure 2: Architecture of self-supervised network with denoising encoder-decoder architecture (illustrated by green box) and classifier output (illustrated by orange box). The input sample is added with additive white Gaussian noise (AWGN).
• Figure 3: Performance of our model on training and testing dataset. Training accuracy is in green color and testing accuracy is in blue.
• Figure 4: Mean activation maps of RBF neurons of three branches. The blue, green and red colors indicate the mean activations of the RBF layers when clean samples are given. Colors black, brown and yellow indicate mean activations under adversarial sample.