Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Securing Large Language Models: Addressing Bias, Misinformation, and Prompt Attacks

Benji Peng, Keyu Chen, Ming Li, Pohsun Feng, Ziqian Bi, Junyu Liu, Xinyuan Song, Qian Niu

TL;DR

The paper surveys security challenges in large language models, focusing on misinformation, bias, content-detection, and adversarial prompt attacks. It comprehensively reviews detection and mitigation strategies, including fact-checking tools, retrieval-augmented generation, and watermarking, along with red-teaming approaches. It examines bias evaluation and mitigation across data, training, and post-processing stages, plus model-editing techniques. Finally, it discusses vulnerabilities such as jailbreaking and prompt injection, and outlines directions for robust defenses and standardized evaluation to advance LLM security in real-world deployments.

Abstract

Large Language Models (LLMs) demonstrate impressive capabilities across various fields, yet their increasing use raises critical security concerns. This article reviews recent literature addressing key issues in LLM security, with a focus on accuracy, bias, content detection, and vulnerability to attacks. Issues related to inaccurate or misleading outputs from LLMs is discussed, with emphasis on the implementation from fact-checking methodologies to enhance response reliability. Inherent biases within LLMs are critically examined through diverse evaluation techniques, including controlled input studies and red teaming exercises. A comprehensive analysis of bias mitigation strategies is presented, including approaches from pre-processing interventions to in-training adjustments and post-processing refinements. The article also probes the complexity of distinguishing LLM-generated content from human-produced text, introducing detection mechanisms like DetectGPT and watermarking techniques while noting the limitations of machine learning enabled classifiers under intricate circumstances. Moreover, LLM vulnerabilities, including jailbreak attacks and prompt injection exploits, are analyzed by looking into different case studies and large-scale competitions like HackAPrompt. This review is concluded by retrospecting defense mechanisms to safeguard LLMs, accentuating the need for more extensive research into the LLM security field.

Securing Large Language Models: Addressing Bias, Misinformation, and Prompt Attacks

TL;DR

The paper surveys security challenges in large language models, focusing on misinformation, bias, content-detection, and adversarial prompt attacks. It comprehensively reviews detection and mitigation strategies, including fact-checking tools, retrieval-augmented generation, and watermarking, along with red-teaming approaches. It examines bias evaluation and mitigation across data, training, and post-processing stages, plus model-editing techniques. Finally, it discusses vulnerabilities such as jailbreaking and prompt injection, and outlines directions for robust defenses and standardized evaluation to advance LLM security in real-world deployments.

Abstract

Large Language Models (LLMs) demonstrate impressive capabilities across various fields, yet their increasing use raises critical security concerns. This article reviews recent literature addressing key issues in LLM security, with a focus on accuracy, bias, content detection, and vulnerability to attacks. Issues related to inaccurate or misleading outputs from LLMs is discussed, with emphasis on the implementation from fact-checking methodologies to enhance response reliability. Inherent biases within LLMs are critically examined through diverse evaluation techniques, including controlled input studies and red teaming exercises. A comprehensive analysis of bias mitigation strategies is presented, including approaches from pre-processing interventions to in-training adjustments and post-processing refinements. The article also probes the complexity of distinguishing LLM-generated content from human-produced text, introducing detection mechanisms like DetectGPT and watermarking techniques while noting the limitations of machine learning enabled classifiers under intricate circumstances. Moreover, LLM vulnerabilities, including jailbreak attacks and prompt injection exploits, are analyzed by looking into different case studies and large-scale competitions like HackAPrompt. This review is concluded by retrospecting defense mechanisms to safeguard LLMs, accentuating the need for more extensive research into the LLM security field.
Paper Structure (14 sections, 2 equations, 3 figures)

This paper contains 14 sections, 2 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Detecting Hallucination
  3. Improve Output Accuracy
  4. Built-in Bias in LLMs
  5. Bias Mitigation Strategies
  6. Detecting LLM-Generated Content
  7. Metric-Based Approaches
  8. Model-Based Approaches
  9. Watermarking and Embedded Signal Approaches
  10. Additonal Challenges
  11. Jailbreaking and Prompt Injection in Large Language Models
  12. Jailbreaking: Exploiting LLM Vulnerabilities
  13. Prompt Injection: Exploiting LLM Input Mechanisms
  14. Future Directions

Figures (3)

  • Figure 1: GPT4o starts to hallucinate when given incorrect information (text in red) by user
  • Figure 2: Common causes and mitigation strategies to hallucination.
  • Figure 3: An attempt to make GPT4o lethargic using prompt injection.