Systems and methods for providing a fraud model for proactively preventing attacks on machine learning models

Abstract

A device may perform adversarial training of a fraud model based on a primary model and with different training data than used to train the primary model, and may receive, from a user, a request for the primary model to generate a prediction. The device may cause the request to be processed by the primary model for generation of a primary model prediction, and may process the request, with the fraud model, to generate a fraud model prediction. The device may perform a comparison analysis of the primary model prediction and the fraud model prediction to generate results, and may determine metadata and relationships based on the primary model prediction, the fraud model prediction, and user behavior patterns. The device may calculate a risk score for the primary model prediction based on the results, the metadata, and the relationships, and may perform actions based on the risk score.