Evaluating Perceptual Hashing Algorithms in Detecting Image Manipulation Over Social Media Platforms

Abstract

Perceptual hash is a fingerprint of features of multimedia content. Compared with crypto hash, perceptual hash shows many advantages when defending against image-based fake news attacks in terms of detecting deliberate image manipulation while still tolerating normal format or resolution changes conducted on user-uploaded images by content-hosting providers such as social media platforms. Previous research into perceptual hash has studied general image manipulation without considering legitimate image transformation by social media platforms. This paper evaluates and analyzes six state-of-the-art perceptual hash algorithms for detecting image manipulation over two major social media platforms: Facebook and Twitter. Our real-world image evaluation shows differences in the two platforms’ image processing and how the six algorithms perform in detecting image manipulation over these platforms. We also present a new approach to finding the optimal detection threshold for each perceptual hash algorithm in distinguishing the platform’s standard image processing from deliberate image manipulation.