A Wi-Fi-Based Authentication Scheme for Voice-Enabled Large Language Model Consumer Electronics Devices

Large Language Models (LLM) agent has proved the remarkable capabilities in transitioning from research projects to widespread applications, which is widely equipped in consumer devices in various domains of technology, business, and society. To improve the user’s experience, LLM agent in consumer electronic devices introduces more inputting manners except for the text input, like the voice interacting. However, it still faces the threats of voice spoofing attacks, such as the replay attack, and the identity spoofing attack. In this paper, we present a Wi-Fi signal based authentication scheme named as LLM-WiLi, which is the first wireless based identification work for LLM. The basic insight of this work is to distinguish the spoofed voice command and the user’s specific identity by analyzing the corresponding wireless signals. To address this, LLM-WiLi establishes a theoretical model to select the wake-up words, followed by the pre-processing and feature extraction module. We propose the temporal feature, frequency feature, and the user mouth motion profile to depict the user’s identity information. Finally, LLM-WiLi exploits the supervised fine-tuning schemes by LLM agents to predict the user’s identity. Experimental results demonstrate LLM-WiLi can achieve an accuracy of 96.4% and 79.9% for defending replay attacks and realizing user identification for 0.29 seconds per trial.