MultiGraSCCo: A Multilingual Anonymization Benchmark with Annotations of Personal Identifiers

A research team led by Ibrahim Baroud has introduced MultiGraSCCo, a groundbreaking multilingual benchmark designed to advance AI-powered anonymization systems. The dataset contains over 2,500 carefully annotated personal identifiers across ten languages, addressing a critical bottleneck in healthcare AI development: accessing sensitive patient data while complying with strict privacy regulations like GDPR and HIPAA. By using synthetic data generation combined with neural machine translation, the researchers created culturally appropriate translations that preserve annotation integrity while rendering names and locations contextually relevant for each target language.

Medical professionals validated the translations for both general quality and specific handling of personal information, confirming the benchmark's utility for real-world applications. The methodology enables researchers to train and test anonymization models without legal complications associated with real patient data. This approach is particularly valuable for low-resource languages, where validated medical data is scarce but privacy requirements remain stringent.

Beyond model development, MultiGraSCCo serves multiple purposes including training human annotators, validating annotations across institutions, and improving automatic personal information detection systems. The team has made both the benchmark and annotation guidelines publicly available, creating a standardized resource that could accelerate privacy-preserving AI research globally. This work represents a significant step toward enabling safe data sharing while maintaining compliance with evolving privacy frameworks.