Healthcare AI Systems Face Data-Poisoning Risks
A 2026 study by Abtahi et al. in Journal of Medical Internet Research synthesizes 41 security studies (2019–2025) to analyze data-poisoning threats across healthcare AI architectures, constructing eight technical attack scenarios spanning CNNs, LLMs, federated learning, resource-allocation systems, and supply chains. It finds attackers with 100–500 poisoned samples can achieve ≥60% success and detection commonly delays 6–12 months, recommending ensemble monitoring, adversarial testing, auditable privacy-preserving mechanisms, and strengthened governance.
Key Points
- 1Demonstrates attackers can compromise models with 100–500 poisoned samples achieving ≥60% success.
- 2Shows attack success depends on absolute poisoned sample count, not dataset proportion, undermining dilution assumption.
- 3Recommends multilayered defenses: ensemble disagreement monitoring, adversarial testing, auditable privacy-preserving governance.
Scoring Rationale
Comprehensive, peer-reviewed synthesis with practical defenses; limited novelty beyond consolidating prior studies and lacking empirical new attacks.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems