Adversarial Attacks on Large Language Models in Medicine.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning Security
• Natural Language Processing Applications

Background:

• Large Language Models (LLMs) show promise for advancing healthcare diagnostics, treatment recommendations, and patient care.
• The deployment of LLMs in sensitive medical contexts is threatened by their vulnerability to adversarial attacks, which could lead to adverse patient outcomes.
• Ensuring the safety and reliability of LLMs in healthcare requires addressing their susceptibility to manipulation.

Purpose of the Study:

• To investigate the vulnerability of both open-source and proprietary LLMs to adversarial attacks in three distinct medical tasks.
• To assess the impact of data poisoning on LLM performance and internal model parameters within medical applications.
• To identify potential methods for detecting and mitigating adversarial attacks on healthcare LLMs.

Main Methods:

• Utilizing real-world patient data to simulate adversarial attacks on LLMs.
• Evaluating LLM performance on medical benchmarks before and after the introduction of poisoned data.
• Analyzing changes in fine-tuned model weights to detect evidence of malicious manipulation.

Main Results:

• Both open-source and proprietary LLMs demonstrated vulnerability to adversarial attacks across multiple medical tasks.
• The integration of poisoned data did not significantly degrade overall LLM performance on standard medical benchmarks.
• Noticeable shifts in fine-tuned model weights were observed, indicating a potential indicator of adversarial manipulation.

Conclusions:

• LLMs in healthcare are susceptible to adversarial attacks, necessitating robust security measures.
• Analysis of model weight shifts presents a promising avenue for detecting and countering attacks on medical LLMs.
• Developing effective defensive mechanisms is crucial for the safe and reliable deployment of LLMs in healthcare settings.