Adversarial Text Perturbation Generation and Analysis

With the evolution of applications of text generations in social networks, the genuineness of such text is questioned. Machine learning language based models such as GPT now can generate responses to complex questions which can be hardly distinguished from human-generated alternatives. In this scope, we utilized text-mutation to evaluate different text-based mutation operators that can be easily created and their impact on the output generated text. Our goal is to evaluate how can machine learning models distinguish those mutated versions from original versions. We reported results of several text-based mutation operators. We evaluated only a few examples of mutation operators and our goal is to eventually create a much larger list of operators. Those mutation operators can be used to distinguish human from machine-generated text.