Adversarial Attacks on IoT Systems Leveraging Large Language Models
Weijie Shan, Teng Long, Zhangbing Zhou · 2024
Internet of Things (IoT) devices have revolutionized various sectors by providing enhanced connectivity and automated services. However, the integration of these devices into complex networks has introduced significant security challenges. Concurrently, advancements in Language Model (LLM) technologies, exemplified by large-scale models like GPT-4, have given rise to sophisticated cyber-attack strategies. With the widespread use of various large language models, these models may generate a large amount of offensive and socially detrimental content during their use. In response, the primary focus of model developers is to adjust alignment models to prevent the generation of harmful content. Meanwhile, adversarial attackers focus on achieving “jailbreaks” to bypass the alignment measures of model developers, thereby generating harmful content and testing the effectiveness of model alignment. The main attack methods currently include prompt injection, unsafe output handling, denial of service, and general adversarial attacks. We selected the Llama-2-7b large language model as the experimental test model and constructed IoT related issues to achieve the generation of harmful content and desired responses.