This paper analyses data poisoning vulnerabilities in healthcare AI systems, showing that a small number of malicious samples can compromise models across architectures, highlighting risks for safety, robustness, and clinical reliability.
Scientific paper - Data Poisoning Vulnerabilities Across Healthcare AI Architectures: A Security Threat Analysis
Anno 2025
This paper looks at how healthcare AI systems can be vulnerable to data poisoning, showing that even a small number of manipulated samples can affect model behaviour. It highlights important gaps in how we currently evaluate AI, especially around robustness and real-world safety, and points out how complex data infrastructures and privacy regulations can make these issues harder to detect.
Author of the paper: Farhad Abtahi et al.
Publisher or journal of publication: arXiv
The paper is available at the following link.
Christine Kakalou, CERTH
Pubblicato il: Giovedì, 01 Maggio 2025 - Ultima modifica: Mercoledì, 06 Maggio 2026
