Exploiting Missing Value Patterns for a Backdoor Attack on Machine Learning Models of Electronic Health Records: Development and Validation Study.

BACKGROUND

A backdoor attack controls the output of a machine learning model in 2 stages. First, the attacker poisons the training data set, introducing a back door into the victim's trained model. Second, during test time, the attacker adds an imperceptible pattern called a trigger to the input values, which forces the victim's model to output the attacker's intended values instead of true predictions or decisions. While backdoor attacks pose a serious threat to the reliability of machine learning-based medical diagnostics, existing backdoor attacks that directly change the input values are detectable relatively easily.

OBJECTIVE

The goal of this study was to propose and study a robust backdoor attack on mortality-prediction machine learning models that use electronic health records. We showed that our backdoor attack grants attackers full control over classification outcomes for safety-critical tasks such as mortality prediction, highlighting the importance of undertaking safe artificial intelligence research in the medical field.

METHODS

We present a trigger generation method based on missing patterns in electronic health record data. Compared to existing approaches, which introduce noise into the medical record, the proposed backdoor attack makes it simple to construct backdoor triggers without prior knowledge. To effectively avoid detection by manual inspectors, we employ variational autoencoders to learn the missing patterns in normal electronic health record data and produce trigger data that appears similar to this data.

RESULTS

We experimented with the proposed backdoor attack on 4 machine learning models (linear regression, multilayer perceptron, long short-term memory, and gated recurrent units) that predict in-hospital mortality using a public electronic health record data set. The results showed that the proposed technique achieved a significant drop in the victim's discrimination performance (reducing the area under the precision-recall curve by at most 0.45), with a low poisoning rate (2%) in the training data set. In addition, the impact of the attack on general classification performance was negligible (it reduced the area under the precision-recall curve by an average of 0.01025), which makes it difficult to detect the presence of poison.

CONCLUSIONS

To the best of our knowledge, this is the first study to propose a backdoor attack that uses missing information from tabular data as a trigger. Through extensive experiments, we demonstrated that our backdoor attack can inflict severe damage on medical machine learning classifiers in practice.