Dynamic threat mitigating of generative artificial intelligence models

Abstract

The disclosure relates to a method and system for dynamically mitigating threats of generative Artificial Intelligence (AI) models. Conventional systems often suffer from inefficiencies due to sequentially applying threat detection checks leading to unnecessary preprocessing and increased computational demands. Additionally, such systems typically focus only on input data, neglecting potential threats in outputs. The disclosed system and method addresses these drawbacks by employing a hierarchical structure of macro and nano classifiers. The system utilizes macro classifiers for broad initial threat categorization followed by specialized nano classifiers for detailed analysis of specific threat subtypes, thereby optimizing processing time and computational resources. The system operates in real time, applying predefined moderation rules to both input and output data to ensure comprehensive threat mitigation. Additionally, continuous telemetry data updates refine nano classifiers and threat identification mechanisms, maintaining high accuracy and adaptability. The disclosed method enhances safety efficiency and reliability of generative AI models.