A Trackable Privacy-Enhanced Machine Learning Classifier
Yun Hu, Chunguo Li, Aiqun Hu, Shunyu Wang · 2022
Nowadays, the privacy protection of machine learning has been taken more and more attentions. Researchers are expected to protect the privacy of training data sets while ensuring the accuracy of machine learning results. This paper attempts to study the privacy protection of machine learning from another direction. A detailed traceable privacy-enhanced machine learning classifier is proposed which combines differential privacy and digital fingerprinting and introduces them together into a classifier algorithm. It can track training devices on the basis of providing more accurate classification results and protecting the privacy of training set. Particularly, privacy protection is achieved by training specific data through the classifier based on differential privacy. And the encoding of the training device identification information is used to specify the location of these specific data. It is the key to the tracking function of the classifier. We use real data to simulate the classifier. Combined with the simulation results, the feasibility, privacy and traceability of the classifier are analyzed. The results show that our classifier scheme can realize tracking training device while satisfying the privacy protection of training set.