Mitigating Query-based Neural Network Fingerprinting via Data Augmentation

Protecting the intellectual property (IP) of deep neural network (DNN) models becomes essential and urgent with rapidly increasing cost DNN training. Fingerprinting is one promising IP protection method that queries suspicious specific fingerprint samples to infer verify by comparing predictions pre-defined labels. Based on utilizing unique features target models, various fingerprinting methods are proposed effectively remotely a meager false-positive ratio. In this paper, we propose novel attack mitigate query-based based data augmentation methods. We randomized transformation input significantly mislead samples’ prediction compromise verification. Then, our can keep model utility an acceptable accuracy drop in data-free scenario (i.e. without any samples) or achieve much higher precision data-limited small number same distribution). An intensive evaluation three well-known structures datasets shows five top-tier conferences.