Adaptive Sampling Technique for Building Knockoff Text Sentiment Models

Commercial machine learning models often require substantial investments of time, money, and expertise, leading organizations to keep critical components such as model architecture, hyperparameters, and training data confidential. This creates black-box models, where users can see the input and output but have no insight into the internal workings. As a result, there has been growing interest in creating knockoff models that replicate the behavior of these black-box models without direct access to their internal details. While much of the current research has focused on black-box attacks in computer vision, there has been relatively little exploration of these attacks in natural language processing, particularly for tasks like text sentiment analysis. In this paper, we address this gap by proposing an adaptive sampling technique that utilizes reinforcement learning to generate more informative training datasets for building knockoff text sentiment models. Unlike traditional random sampling methods, our approach interacts with the victim model during data selection, continuously refining the knockoff model by focusing on data points that better capture the behavior of the original model. Experimental results show that this adaptive sampling technique significantly improves the accuracy and behavioral fidelity of the knockoff models, closely mimicking the performance of the victim models.