Adversarial attack detection in voice-based Parkinson’s disease diagnosis using machine learning

Parkinson’s disease (PD) is a globally prevalent progressive neurodegenerative disorder. Early diagnosis is essential to slow disease progression and improve patients’ quality of life. Machine learning (ML)-based PD diagnostic approaches using voice biomarkers have shown particularly encouraging results; however, ML models are prone to various cyberattacks, including adversarial examples that manipulate input data to mislead model predictions. In this work, we propose a secure ML-based PD diagnostic framework for voice recordings that includes both adversarial example simulation and attack detection components. Experiments were conducted using the publicly available UCI Parkinson’s Disease Dataset with multimodal classifiers based on Support Vector Machine (SVM), Random Forest (RF), AdaBoost, and Light Gradient Boosting Machine (LightGBM). The classifiers achieved accuracies of 87% for SVM, 88% for RF, 90% for AdaBoost, and 91% for LightGBM when tested on benign samples from the original dataset. A detailed analysis and comparison of their performance under adversarial perturbations designed as imperceptible worst-case attacks to reduce model accuracy, along with the proposed second-stage classifier for detecting adversarial inputs, revealed that, without proper countermeasures, ML-based decision-making systems can be highly vulnerable to cyber manipulation. Such attacks resulted in significant performance degradation, reducing the accuracy to 86%, 89%, 84%, and 86%, respectively.