Ling Liu

Tiansheng Huang, Sihao Hu, Fatih Ilhan et al. · gatech

Recent research shows that Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- models lose their safety alignment ability after fine-tuning on a few harmful samples. For risk mitigation, a guardrail is typically used to filter out harmful samples before fine-tuning. By designing a new red-teaming method, we in this paper show that purely relying on the moderation guardrail for data filtration is not reliable. Our proposed attack method, dubbed Virus, easily bypasses the guardrail moderation by slightly modifying the harmful data. Experimental results show that the harmful data optimized by Virus is not detectable by the guardrail with up to 100\% leakage ratio, and can simultaneously achieve superior attack performance. Finally, the key message we want to convey through this paper is that: \textbf{it is reckless to consider guardrail moderation as a clutch at straws towards harmful fine-tuning attack}, as it cannot solve the inherent safety issue of the pre-trained LLMs. Our code is available at https://github.com/git-disl/Virus

Xinxin Fan, Ling Liu, Rui Zhang et al.

Decentralized trust management is used as a referral benchmark for assisting decision making by human or intelligence machines in open collaborative systems. During any given period of time, each participant may only interact with a few of other participants. Simply relying on direct trust may frequently resort to random team formation. Thus, trust aggregation becomes critical. It can leverage decentralized trust management to learn about indirect trust of every participant based on past transaction experiences. This paper presents alternative designs of decentralized trust management and their efficiency and robustness from three perspectives. First, we study the risk factors and adverse effects of six common threat models. Second, we review the representative trust aggregation models and trust metrics. Third, we present an in-depth analysis and comparison of these reference trust aggregation methods with respect to effectiveness and robustness. We show our comparative study results through formal analysis and experimental evaluation. This comprehensive study advances the understanding of adverse effects of present and future threats and the robustness of different trust metrics. It may also serve as a guideline for research and development of next generation trust aggregation algorithms and services in the anticipation of risk factors and mischievous threats.