Back

Overlooked Trustworthiness of Explainability in Medical AI

Zhang, J.; Chao, H.; Kalra, M. K.; Wang, G.; Yan, P.

2021-12-24 radiology and imaging
10.1101/2021.12.23.21268289 medRxiv
Show abstract

While various methods have been proposed to explain AI models, the trustworthiness of the generated explanation received little examination. This paper reveals that such explanations could be vulnerable to subtle perturbations on the input and generate misleading results. On the public CheXpert dataset, we demonstrate that specially designed adversarial perturbations can easily tamper saliency maps towards the desired explanations while preserving the original model predictions. AI researchers, practitioners, and authoritative agencies in the medical domain should use caution when explaining AI models because such an explanation could be irrelevant, misleading, and even adversarially manipulated without changing the model output. AI researchers, practitioners, and authoritative agencies in the medical domain should use caution when explaining AI models because such an explanation could be irrelevant, misleading, and even adversarially manipulated without changing the model output.

Matching journals

The top 6 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.