Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) are key advancements in artificial intelligence (AI) capable of generating text, interpreting images, and understanding complex multimodal inputs, mimicking human intelligence. However, concerns arise due to their potential misuse and vulnerabilities to jailbreak attacks, where malicious inputs trick the models into generating harmful or objectionable content.
Securing AI models against such attacks requires identifying and mitigating any vulnerabilities. This is a challenging task as it requires a thorough understanding of how AI models can be manipulated. Researchers have developed various testing and evaluation methods to probe the defenses of LLMs and MLLMs, such as altering textual inputs and introducing visual perturbations, to assess models’ adherence to safety protocols under various attack scenarios.
Researchers from various institutions, including LMU Munich, University of Oxford, Siemens AG, Munich Center for Machine Learning (MCML), and Wuhan University, have proposed a comprehensive framework for evaluating AI model robustness. The framework involves a dataset comprising 1,445 harmful questions across 11 distinct safety policies. The study applied extensive red-teaming approach, testing the resilience of 11 different LLMs and MLLMs, including proprietary models like GPT-4 and GPT-4V, and open-source models. The objective was to uncover weaknesses in the models’ defenses, providing insights that can be used to strengthen them against potential attacks.
The study’s methodology involves both hand-crafted and automatic jailbreak methods to simulate a range of attack vectors. The objective was evaluating the models’ maintenance of safety protocols despite sophisticated manipulation tactics. The findings revealed that GPT-4 and GPT-4V exhibited superior robustness compared to their open-source counterparts, demonstrating their resistance to jailbreak attempts effectively. Notably, among the open-source models, Llama2 and Qwen-VL-Chat demonstrated robustness, with Llama2 even surpassing GPT-4 in some scenarios.
This research provides substantial insights into the ongoing discourse on AI safety, presenting a nuanced analysis of LLMs and MLLMs’ vulnerability to jailbreak attacks. Via systematic performance evaluations of various models against an extensive range of attacks, the study identifies current weaknesses while setting a benchmark for future model improvements. The data-driven approach sets a new standard for assessing AI model security via the incorporation of diverse harmful questions and comprehensive red-teaming techniques.
In conclusion, the study highlights the significant security risks posed by the vulnerability of LLMs and MLLMs to jailbreak attacks. These results, achieved through a robust evaluation framework involving an extensive harmful query dataset and rigorous red-teaming techniques, provide a comprehensive assessment of AI model security. Reportedly, proprietary models such as GPT-4 and GPT-4V demonstrated exceptional resilience against these attacks, outclassing their open-source counterparts.
