Poisoning Attacks via Generative Adversarial Text to Image Synthesis

Abstract

A poisoning attack is where the adversary can inject a small fraction of poisoning instances into the training data used to train a machine learning model to compromise the performance. Poison attacks can significantly affect the learning process and performance as the model is trained on incorrect data. We have seen many works on data poisoning over the years, but it is limited to few deep learning networks. In this work, we introduce a novel approach by leveraging Generative Adversarial Text to Image Synthesis to create poison attacks against machine learning classifiers. Our approach has three components, which are the generator, discriminator, and the target classifier. We performed an extensive experimental evaluation that proves our attack's efficiency to compromise machine learning classifiers including deep networks.