Evaluating Machine Learning Models for Handwriting Recognition-based Systems under Local Differential Privacy

Handwriting recognition, a pervasive assistive technology for a variety of consumer electronics and Internet of Things (IoT) systems does not provide meaningful privacy guarantees. Hacking of a device, malware, and compromise of handwriting recognition-enabled application's cloud service can lead to leakage of sensitive personal information with serious consequences. In this paper, we propose an algorithm to deform a user's original handwriting by adding adjustable statistical noise before sharing it with a specific application to achieve local differential privacy (LDP). To assess and quantify the algorithm's impact on recognition, we generate a shape-deformed test set by applying the algorithm to the MNIST dataset and implement a visualization tool to demonstrate the impact of the proposed algorithm. Then, we train a convolutional neural network, Naive Bayes, random forest, support vector machine, and decision tree classifiers on the MNIST dataset and evaluate them with the privatized test set. Our experiment reveals that without tuning the CNN for privatized images, it is possible to achieve around 80% accuracy in local differential privacy settings, which lays the ground for further research in this direction.