Fighting AI-Camera-Captured Image Manipulation with AI-Enabled Solutions

Description

The era of artificial intelligence (AI) has witnessed increasing volumes of high-quality image content through smart photography and popular image applications. However, the advance of AI also makes high-quality image manipulation (i.e., image splicing, inpainting, copy-move) much easier, which threatens image integrity and can lead to serious security and ethical issues. Harmful consequences include public disinformation and damaged personal reputation from pornographic images created through face-swapping technologies. Therefore, there is an increasing need for image manipulation detection tools to counter such threats. Previous research on forensics and image manipulation detection has mainly focused on how conventional digital cameras introduce intrinsic image regularities when their image signal processor (ISP) processes photographs. Image manipulation affects those regularities and thus leaves digital traces. However, the advent of new-generation AI cameras (e.g., smartphones’ built-in cameras using AI technologies) poses a major challenge because they use a smart ISP based on deep neural networks. Owing to the non-linear structure of deep neural networks, image regularities introduced by AI cameras can be much more complex than those introduced by conventional cameras, rendering existing image manipulation detection techniques ineffective. Furthermore, an adversary can launch an anti-forensics attack by using an AI camera to recapture a manipulated image in high quality to hide manipulation traces. Hence, it is urgent to develop advanced techniques to detect manipulated AI-camera-captured images, and to prevent and respond to anti-forensic attacks. In this proposal, we aim to create a generalized and robust system to detect manipulation of AI-camera-captured images. Importantly, our system will be powered by AI and based on promising results from our preliminary experimental study. Our research tasks for the system design are as follows: (1) to design a framework for digital fingerprint extraction that accounts for key neural network features of smart ISPs in AI cameras; (2) to design an algorithm to detect manipulation of AI-camera-captured images that is generalizable to various AI camera models for real-world application; and (3) to design an algorithm to counter anti-forensics attacks and further improve the robustness of our image manipulation detection system. By harnessing novel AI-enabled solutions, our proposal aims to fight the threat to image integrity caused by AI cameras in the new generation of smart photography. We expect to advance the research area of digital image forensics, while also benefiting society in many areas including security, law, journalism, social media, business, and research integrity.