Model Extraction Attacks on Arabic BERT-Based APIs

Hassan Abbelkarim, Mohammed Eltahir, Khalid Elmadani, Anas Showk

Keywords: Model Extraction, Protection, Arabic NLP

TL;DR: We study Model Extraction Attacks on Arabic BERT Base models and propose Protection strategy

Abstract: In this paper, we study the feasibility of performing Model Extraction attacks on Arabic BERT-based APIs. In our experiments, we try to perform these attacks under different scenarios and observe the accuracy of the extracted model against the victim model. We then propose a method for protecting against these types of attacks by introducing noise, in the form of pre-training the victim model for more epochs on non-public generic data. Our results show that this strategy better secures the victim model from such attacks.