Benchmarking Scientific Image Forgery Detectors.

Area of Science:

• Computer Vision
• Digital Forensics
• Scientific Integrity

Background:

• Lack of accessible datasets hinders the development and evaluation of scientific image forensic techniques.
• Data sensitivity and legal hurdles restrict the use of real-world cases for creating forensic benchmarks.
• Limited understanding of automatic image analysis tool capabilities may lead to a false sense of data integrity.

Purpose of the Study:

• To address the bottleneck in scientific image integrity research by creating an open-source library and a large-scale benchmark dataset.
• To reproduce common image forgery operations (duplication, retouching, cleaning) for synthetic dataset generation.
• To evaluate the performance of state-of-the-art copy-move detection methods on scientific images.

Main Methods:

• Developed an extendable open-source algorithm library to simulate common image forgery operations.
• Created a large scientific forgery image benchmark comprising 39,423 synthetically doctored images with enriched ground truth.
• Collected source images from creative commons, ensuring they had no pre-existing integrity issues.
• Evaluated copy-move detection methods using a novel metric for consistent match detection between source and copied regions.

Main Results:

• The created benchmark dataset contains 39,423 synthetically generated scientific forgery images.
• All evaluated state-of-the-art copy-move detection methods demonstrated low performance on the proposed dataset.
• A new metric was introduced to assert consistent match detection for copy-move forgery evaluation.

Conclusions:

• Scientific images may require specialized copy-move detectors due to the inadequacy of current general methods.
• The open-source library and benchmark dataset facilitate further research in scientific image forensics.
• The findings highlight the need for improved tools to ensure the integrity of scientific visual data.