Adversarial Fragility and Language Vulnerability in Clinical AI

A framework for auditing clinical AI safety. The work demonstrates two vulnerabilities: imperceptible perturbations that collapse diagnostic accuracy from 89.3% to 62.0% on chest X-ray classification, and cross-lingual drift where models drop measurably across English, Nigerian Pidgin, and Yoruba-inflected English. Argues for adversarially robust, multilingual architectures in low-resource clinical deployments.

Invisible perturbations collapse the diagnosis. Cross-lingual drift erases the patient.