Exploiting Vulnerabilities of Machine Learning Models on Medical Text via Generative Adversarial Attacks

Artificial Intelligence (AI) technology has undergone significant advancements, driving its adoption across various fields. AI, especially Machine Learning (ML) adoption has also grown widely in the medical field, which naturally requires actions with high precision. Challenges arise when AI-based predictive models prove to be vulnerable to adversarial attacks. These attacks use perturbed data (modified data), which is unnoticeable to humans but can significantly alter prediction results. This paper aims to implement the TextFooler technique to deceive predictive models on medical text against adversarial attacks. The experiment reveals that three ML models developed using popular approachs i.e., transformer-based model based on Bidirectional Encoder Representations from Transformers (BERT), Stack Classifier methods and traditional machine learning algorithms achieved the same classification accuracy of 99.98%. Among the three models, BERT is the most affected by adversarial attacks, with an attack success rate of 76.8%, followed by traditional machine learning methods and the stack classifier, with success rates of 28.73% and 5.21%, respectively. These findings indicate that although the advanced ML model BERT demonstrates good performance, it is highly vulnerable to adversarial attacks. Therefore, there is an urgency to develop predictive models that are robust and secure against potential attacks.