Using binary hash tree-based encryption to secure a deep learning model and generated images for social media applications

Abstract

Deep learning (DL) plays a vital role in identifying critical features and patterns in digital images. Deep learning models and generated records, particularly digital images, are highly effective in media and other applications but pose privacy and security challenges. For example, healthcare professionals must understand how Artificial Intelligence (AI) makes decisions to trust and fully incorporate its findings into medical practice. This research addresses the security and privacy challenges associated with a DL model and generated records for social media applications. In this work, we propose a binary hash tree-based encryption that encrypts a customised model and generated images to minimise data leakage. The proposed method includes three parts. First is a customised autoencoder that minimises the size of digital images ensuring the security of generated images with a Henon chaotic map and ephemeral keys derived from a binary hash tree (BHT) for encryption in a Galois field (GF). Further, we encrypt the fewest possible weight parameters of the customised model with the same ephemeral key to preserve privacy. By doing so, our method reduces data leakage and further improves the model security at the same time. Extensive experiments reveal that the proposed method is more secure against attacks than state-of-the-art methods, which could be helpful in media and several other applications. To the best of our knowledge, we are the first to explore a secure system that protects both the model and the generated media at the same time using an encryption technique.