Image Manipulation Dataset

Dataset Composition

The DF2023 training dataset is extensive, consisting of one million high-quality images, each categorized under four predominant forgery techniques. These categories reflect the most common and impactful forms of image manipulation currently observed in the digital landscape:

• Splicing (400,000 images): This category involves the merging of elements from multiple images to create a single, composite image. The splicing method is frequently used to fabricate events or scenarios that never occurred.
• Copy-Move (300,000 images): In this type of manipulation, a portion of an image is copied and pasted elsewhere within the same image. This technique is often employed to conceal or duplicate specific features.
• Enhancement (200,000 images): This set contains images that have been digitally enhanced to alter their visual impact, such as modifying brightness, contrast, or colors to misrepresent the original content.
• Removal (100,000 images): This category includes images where objects or people have been removed, leaving a manipulated scene that distorts the original context.

Data Quality and Standards

Each image in the DF2023 dataset has undergone rigorous quality checks to ensure its relevance and utility for high-stakes research. The dataset is formatted to be compatible with various machine learning frameworks, and it includes detailed metadata to facilitate ease of use. Additionally, the dataset adheres to ethical standards, with all manipulations being clearly documented to prevent misuse.