Hyperdimensional Multimedia Perception and Frontier Security

Background

In the era of easy-to-use editing software, distinguishing genuine media from manipulated ones has become a major challenge. Our research focuses on developing intelligent algorithms that can automatically detect and localize tampered regions in images and videos, especially those modified by popular editing software. By combining digital forensics and deep learning, we aim to uncover invisible editing traces, enhance media authenticity, and support a more trustworthy digital ecosystem.

Common Types of Media Tampering

Digital image and video tampering typically falls into three main categories: copy-move, splicing and removal. Each leaving unique forensic traces that our research aims to detect and localize.

Copy-move forgery copies a region from an image and pastes it within the same image to hide or duplicate elements. This technique preserves image statistics, making it harder to detect. Our models leverage self-similarity analysis and keypoint feature matching to locate duplicated regions.

Splicing combines regions from two or more different images into one composite image, often to insert or replace objects. Common examples include adding people or objects to a scene. Our algorithms detect inconsistencies in lighting, color tones, and sensor patterns to reveal the manipulated boundaries.

Removal forgery erases unwanted objects or persons and fills the gap using inpainting or content-aware fill tools. Modern editing software can make removals almost invisible to the human eye. Our detection methods identify texture inconsistencies and structural discontinuities to expose such manipulations.