Differentially Private Machine Learning Model against Model Extraction Attack

Machine learning model is vulnerable to model extraction attacks since the attackers can send plenty of queries to infer the hyperparameters of the machine learning model thus stealing confidential information of the learning models. Therefore, there is a urgent need to defend against such an attack. Differential privacy is a promising technique to protect the valuable information. We propose a differential privacy-based method applied in the linear neural network to obfuscate the output of the machine learning model. The security and utility issue of injecting a noise layer to the linear neural network is mathematically analyzed. The experiment results show that our proposed method can lower the attacker's extraction rate while keeping high utility.