Adversarial training to minimize data poisoning attacks

Abstract

The techniques disclosed herein enable systems to train machine learning models using benign augmentation to enabled resistance various data poisoning attacks. This is achieved by first training a machine learning model using an initial dataset that is trustworthy and originates from a known source. The initial dataset is then modified to include known attack triggers such as syntactic paraphrasing to generate an augmented dataset. The augmented dataset is then used to train a robust machine learning model based using the initially trained machine learning model. The resultant robust machine learning model is then enabled to detect and resist attacks captured by the augmented dataset. The robust machine learning model can be retrained using an untrusted dataset that includes various compromised inputs in conjunction with the augmented dataset. Retraining results in an updated robust machine learning model that can learn and resist various data poisoning attacks on the fly.