# Adversarial Attacks and Defenses in Large Language Models: Old and New Threats

**Authors:** Leo Schwinn, David Dobre, Stephan Günnemann, Gauthier Gidel  
**arXiv:** [2310.19737](https://arxiv.org/abs/2310.19737)  
**Date:** 2023-10-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.19737v1/

## Key points

- Challenges in enhancing the robustness of neural networks
- Overestimation of new defense approaches due to faulty evaluations
- Proposal of prerequisites to enhance the evaluation process and reduce faulty assessments
- Focus on adversarial arms race in natural language processing with closed-source Large Language Models (LLMs)
- Identification of embedding space attacks as a viable threat model for generating malicious content in open-sourced models
- Need for improved defenses against embedding space attacks
- Analysis of a recently proposed defense approach and demonstration of overestimating robustness without tailored best practices for LLMs
- Emphasis on accurate robustness evaluations in neural networks
- Insights into potential threats and challenges associated with large language models
- Importance of implementing effective defense strategies to enhance network robustness in natural language processing applications.

## Summary

In the paper titled "Adversarial Attacks and Defenses in Large Language Models: Old and New Threats," authors Leo Schwinn, David Dobre, Stephan Günnemann, and Gauthier Gidel discuss the challenges associated with enhancing the robustness of neural networks. Despite extensive research over the past decade, this problem remains largely unsolved. One major obstacle in improving robustness is the overestimation of new defense approaches due to faulty evaluations. Flawed assessments of robustness not only slow down research but also provide a false sense of security. To address this issue, the authors propose a set of prerequisites that can enhance the evaluation process and reduce the number of faulty assessments. The authors specifically focus on an impending adversarial arms race in natural language processing, particularly with closed-source Large Language Models (LLMs) like ChatGPT, Google Bard, or Anthropic's Claude. These LLMs pose significant challenges in terms of robustness assessment. Additionally, the authors identify embedding space attacks as another viable threat model for generating malicious content in open-sourced models. This highlights the need for improved defenses against such attacks. To demonstrate their point, the authors analyze a recently proposed defense approach. They show that without specific best practices tailored to LLMs in place, it is easy to overestimate the robustness of a new approach. Overall, this paper emphasizes the importance of accurate robustness evaluations in neural networks and provides insights into potential threats and challenges associated with large language models. By addressing these issues and implementing effective defense strategies, researchers can make significant progress towards enhancing network robustness in natural language processing applications.