Researchers Poison Stolen Data to Sabotage AI Model Accuracy

AURA is a defensive framework that deliberately adulterates proprietary knowledge graphs with structurally plausible but factually incorrect entries to degrade the utility of stolen AI databases for unauthorized users while preserving full fidelity for authorized users via a cryptographic filter. The research reports strong degradation of unauthorized model accuracy (to single-digit percentages in tests), high stealth against anomaly detectors, resilience to sanitization, and minimal overhead for legitimate queries.