SmartHash: Perceptual Hashing for Image Tampering Detection and Authentication

Perceptual hashing algorithms have been used extensively to detect duplicate images, similar images for reverse image search, inappropriate and explicit images, and child sexual abuse (CSAM) images. These algorithms use various techniques to extract perceptual features from an image to create a succinct representation called the hash which is akin to the bio-metric marker of an image. This paper explores the ability of perceptual hashes to determine whether an image is a tampered version of a previously known image, i.e., it was created by object addition or removal from the original known image. In particular, a fast and efficient DCT-based perceptual hashing algorithm called SmartHash is proposed. SmartHash is an extension of a well-known and widely used pHash algorithm. It is evaluated on several publicly available datasets of tampered images and is shown to have high accuracy and precision in detecting such images. Additionally, an in-depth examination of the results is provided to quantify the limitations and hence, operating parameters of any software that integrates SmartHash in its workflow. Comparison with the state of the art Apple's NeuralHash and Microsoft's PhotoDNA is provided in the context of detecting tampered images. This work contributes to the content authentication initiative which is led by Adobe to establish content provenance and authentication across the Internet.