Parkinson Disease (PD) is a progressive nervous system disease that interferes with motor control and speech, and therefore, it is necessary to detect it early but challenging because there are no conclusive lab tests. Clinical assessment is commonly used in the traditional diagnosis and this might be subjective and time consuming. The proposed research is intended to develop a reliable and correct artificial intelligence system of PD detection based on voice biomarkers. The publicly available speech features Kaggle dataset was used. In the preprocessing stage, such techniques as feature scaling and dimensionality reduction were used, six machine learning models were run: LightGBM, CatBoost, Extra Trees, Quadratic Discriminant Analysis (QDA), Gaussian Process Classifier (GPC), and Multi-layer perceptron (MLP). The findings show that GPC and MLP performed better with 95% accuracy and CatBoost was second with 94%. Adversarial attacks (FGSM and PGD) were also proposed to evaluate robustness, and a substantial decrease in performance was found. This was followed by the use of adversarial training that somewhat regained accuracy and enhanced resistance to attacks. The novelty of this work consists in collaboration between the cybersecurity perspectives and medical AI in the detection of PD so that models are not only precise but also robust. This system provides a secure, nonintrusive, deployable telemedicine application. A quick comparison to the state-of-the-art (CNN, LSTM, ensemble models) shows our system similar or better accuracy along with more adversarial robustness.