Adversarial Machine Learning in Healthcare: Risks to AI-Driven Diagnostics and Treatment Plans

Abstract

The rapid integration of artificial intelligence (AI) in healthcare has enhanced diagnostics, predictive analytics, and clinical decision-making. However, AI-driven models, particularly deep learning architectures, remain highly vulnerable to adversarial machine learning (AML) attacks, which can result in misdiagnoses, unsafe treatment recommendations, and compromised patient safety. This study systematically evaluates adversarial risks in medical AI, quantifies their impact on model performance, and assesses the efficacy of defense mechanisms. We analyzed CNNs (medical imaging), RNNs (ECG analysis), and Transformer models (clinical NLP) under FGSM, PGD, and JSMA attacks. Results show that the CNN accuracy of 92% was reduced to 40% under JSMA, ECG-based AI performance dropped by 42% under PGD, and Transformer-based NLP models experienced a 30% decline under FGSM. Defense mechanisms such as randomized smoothing and adversarial training improved accuracy by 15% and 14%, respectively, though at high computational costs (1.8× and 1.5× training overhead). Across five independent trials, all degradations were statistically significant (p< 0.01), and ANOVA with Tukey’s HSD confirmed that randomized smoothing and adversarial training significantly outperformed gradient masking (p< 0.01). These findings demonstrate that medical AI systems are highly susceptible to adversarial manipulation and underscore the necessity of robust, efficient, and regulatory-compliant defenses. Strengthening adversarial resilience is critical to ensuring safe, reliable, and ethically responsible deployment of AI in healthcare.