Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models

Created by
  • Haebom

Author

Runpeng Dai, Run Yang, Fan Zhou, Hongtu Zhu

Outline

This paper analyzes the vulnerability of large-scale language models (LLMs) and vision-language models (VLMs) to identify the cause of their high sensitivity to small changes. We propose a new stability metric, First-order local influence (FI), based on information geometry to quantify the sensitivity of individual parameters and input dimensions (pixel or token embeddings). By analyzing various LLMs and VLMs with parameters ranging from 1.5 billion to 13 billion, we reveal that a small number of parameters or input dimensions with high FI values disproportionately contribute to the vulnerability of the models, and show that mitigating the influence of these vulnerable parameters contributes to improved model performance.

Takeaways, Limitations

•
Takeaways:
◦
We propose a new approach to model vulnerability analysis by introducing a new stability measurement index, FI, based on information geometry.
◦
By identifying parameters and input dimensions that contribute to the model's vulnerability, we propose targeted strategies to improve the model's robustness.
◦
We experimentally demonstrate that mitigating the influence of vulnerable parameters leads to improved model performance.
•
Limitations:
◦
Further research is needed on the generality of FI metrics and their applicability to various model architectures.
◦
Further research and more efficient methodologies are needed to mitigate the effects of vulnerable parameters.
◦
The scale and types of models used in the analysis may be limited, and analysis of a wider variety of models is needed.
👍
Made with Slashpage