An Autoencoder-Based Image Anonymization Scheme for Privacy Enhanced Deep Learning

The development of deep learning (DL) technology is dependent on the availability of large-scale image datasets to train deep neural networks (DNNs) for image classification. However, many raw image datasets contain sensitive identity feature information that prohibit entities from disclosing data due to privacy regulations. For example, an image dataset may include age or gender information that could be used to identify an individual. Furthermore, medical images may include additional disease information that could lead to patient re-identification. To address this problem, we propose an image transformation scheme using a convolutional autoencoder and multi-output classification model for privacy enhanced deep learning. The proposed scheme obfuscates image visual information while retaining useful attribute features that are required for model utility. Additionally, the proposed method enhances privacy by generating encoded images that exclude sensitive identity feature information. First, we train a multi-output convolutional neural network (CNN) to classify identity features and image attributes. Second, we use the pre-trained multi-output classifier for regularization in training a standard convolutional autoencoder to generate obfuscated versions of the original images that exclude identity feature information and preserve attribute features that are useful for classification. Our results on CelebA and Cifar-100 datasets illustrate that the proposed method successfully degrades classification accuracy of sensitive image data while maintaining model utility for non-sensitive data features.