A system and method for programmatically revealing misleading confidence values in Fraud Score is presented to protect artificial intelligence models from adversarial neural networks. The method is used to reduce an adversarial learning neural network model effectiveness. With the score manipulation implemented, the adversary models are shown to systematically become less successful in predicting the true behavior of the Fraud detection artificial intelligence model and what it will flag as fraudulent transactions, thus reducing the true fraud dollars penetrated or taken by adversaries.