The Differential Privacy Framework for Convolutional Neural Network using PATE, ADLM and SGD models
Suwan Tang, Wu Wang · 2023
The use of CNNs raises concerns over privacy, especially when dealing with sensitive data, therefore, differential privacy has emerged to protect the privacy of data providers while allowing data analysis and model training. Differential privacy is a mathematical framework that protects sensitive data by adding random noise to the training data. In this study, we propose a differential privacy framework for CNNs using the PATE(Private Aggregation of Teacher Ensembles), ADLM(Adaptive Laplace Mechanism), and SGD(Stochastic Gradient Descent) models. We experimentally demonstrate that the proposed framework achieves better accuracy while providing strong privacy guarantees in image recognition tasks.