Data Poisoning Vulnerabilities Across Health Care Artificial Intelligence Architectures: Analytical Security...

Area of Science:

• Medical Artificial Intelligence
• Cybersecurity in Healthcare
• AI Model Security

Background:

• Health care AI systems are increasingly integrated into clinical workflows but remain vulnerable to data-poisoning attacks.
• A small number of manipulated training samples can compromise AI models used for diagnosis, documentation, and resource allocation.
• Existing privacy regulations may inadvertently complicate anomaly detection and auditing, limiting visibility into adversarial activity.

Purpose of the Study:

• Provide a comprehensive threat analysis of data poisoning vulnerabilities across major health care AI architectures.
• Identify attack surfaces in clinical AI systems.
• Propose a multilayered defense framework suitable for health care settings.

Main Methods:

• Synthesized empirical findings from 41 security studies (2019-2025) into a health care-specific threat-modeling framework.
• Constructed 8 hypothetical attack scenarios across AI architectures, federated learning, clinical documentation, resource allocation, and supply chains.
• Aligned scenarios with realistic insider-access threat models and current clinical deployment practices.

Main Results:

• Attackers with as few as 100-500 poisoned samples can compromise health care AI systems (≥60% success rate).
• Attack success depends on the absolute number of poisoned samples, challenging assumptions about large datasets providing inherent protection.
• Detection delays range from 6-12 months, potentially years in distributed or privacy-constrained environments; privacy regulations further complicate detection.

Conclusions:

• Health care AI systems face significant security challenges inadequately addressed by current regulations and validation practices.
• A multilayered defense strategy is proposed, combining ensemble disagreement monitoring, adversarial testing, privacy-preserving auditable mechanisms, and strengthened governance.
• Ensuring patient safety may require a shift towards more interpretable and verifiable AI architectures over opaque, high-performance models.