Model Stealing Defense with Hybrid Fuzzy Models: Work-in-Progress

摘要

With increasing applications of Deep Neural Networks (DNNs) to edge computing systems, security issues have received more attentions. Particularly, model stealing attack is one of the biggest challenge to the privacy of models. To defend against model stealing attack, we propose a novel protection architecture with fuzzy models. Each fuzzy model is designed to generate wrong predictions corresponding to a particular category. In addition’ we design a special voting strategy to eliminate the systemic errors, which can destroy the dark knowledge in predictions at the same time. Preliminary experiments show that our method substantially decreases the clone model's accuracy (up to 20%) without loss of inference accuracy for benign users.