Back

Exposing Vulnerabilities in Clinical LLMs Through Data Poisoning Attacks: Case Study in Breast Cancer

Das, A.; Tariq, A.; Batalini, F.; Dhara, B.; Banerjee, I.

2024-03-21 health informatics
10.1101/2024.03.20.24304627 medRxiv
Show abstract

Training Large Language Models (LLMs) with billions of parameters on a dataset and publishing the model for public access is the standard practice currently. Despite their transformative impact on natural language processing, public LLMs present notable vulnerabilities given the source of training data is often web-based or crowdsourced, and hence can be manipulated by perpetrators. We delve into the vulnerabilities of clinical LLMs, particularly BioGPT which is trained on publicly available biomedical literature and clinical notes from MIMIC-III, in the realm of data poisoning attacks. Exploring susceptibility to data poisoning-based attacks on de-identified breast cancer clinical notes, our approach is the first one to assess the extent of such attacks and our findings reveal successful manipulation of LLM outputs. Through this work, we emphasize on the urgency of comprehending these vulnerabilities in LLMs, and encourage the mindful and responsible usage of LLMs in the clinical domain.

Matching journals

The top 7 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.