IEEE VIS 2024 Content: AdversaFlow: Visual Red Teaming for Large Language Models with Multi-Level Adversarial Flow

 Honorable Mention

AdversaFlow: Visual Red Teaming for Large Language Models with Multi-Level Adversarial Flow



Dazhen Deng - Zhejiang University, Ningbo, China

 Chuhan Zhang - Zhejiang University, Hangzhou, China

 Huawei Zheng - Zhejiang University, Hangzhou, China

 Yuwen Pu - Zhejiang University, Hangzhou, China

 Shouling Ji - Zhejiang University, Hangzhou, China

 Yingcai Wu - Zhejiang University, Hangzhou, China

 Screen-reader Accessible PDF

 Room: Bayshore V

2024-10-18T12:30:00ZGMT-0600Change your timezone on the schedule page
2024-10-18T12:30:00Z

Exemplar figure, described by caption below — The interface of AdversaFlow includes a Control Panel (A) to configure model parameters and adjust data sampling, an Embedding View (B) to show the projection of prompts, a Metric Monitor (C) displaying the key performance indicators of the model, an Adversarial Flow to facilitate multi-level exploration of models, an Instance List (E) showing prompt details, and a Flucutaion View (F) for the investigation of token-level uncertainty.

Fast forward

Full Video

Keywords

Visual Analytics for Machine Learning, Artificial Intelligence Security, Large Language Models, Text Visualization

Abstract

Large Language Models (LLMs) are powerful but also raise significant security concerns, particularly regarding the harm they can cause, such as generating fake news that manipulates public opinion on social media and providing responses to unethical activities. Traditional red teaming approaches for identifying AI vulnerabilities rely on manual prompt construction and expertise. This paper introduces AdversaFlow, a novel visual analytics system designed to enhance LLM security against adversarial attacks through human-AI collaboration. AdversaFlow involves adversarial training between a target model and a red model, featuring unique multi-level adversarial flow and fluctuation path visualizations. These features provide insights into adversarial dynamics and LLM robustness, enabling experts to identify and mitigate vulnerabilities effectively. We present quantitative evaluations and case studies validating our system's utility and offering insights for future AI security solutions. Our method can enhance LLM security, supporting downstream scenarios like social media regulation by enabling more effective detection, monitoring, and mitigation of harmful content and behaviors.