Deep Learning-Based Forgery Attack on Document Images

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Digital document images are prevalent in online services.
• Deep learning enables end-to-end text editing in images.
• Existing algorithms struggle with complex characters and backgrounds.

Purpose of the Study:

• To develop a low-cost document forgery algorithm for practical image editing.
• To address limitations of current text editing methods in complex scenarios.
• To create realistic forged documents that evade detection.

Main Methods:

• Disentangling text and background information for clearer supervision data.
• Utilizing text skeleton as auxiliary information for complex structures.
• Incorporating texture continuity and print-and-scan distortions in post-processing.
• Employing a set of network design strategies to mitigate forgery traces.

Main Results:

• Reduced reconstruction error by approximately 2/3 (MSE).
• Improved reconstruction quality by 4 dB (PSNR) and 0.21 (SSIM).
• Visually superior results on complex characters and textures compared to existing methods.
• Successfully fooled existing document authentication systems in practical scenarios.

Conclusions:

• The proposed method offers a significant advancement in document forgery.
• It effectively handles complex visual elements in document images.
• Demonstrates practical applicability in identity document alteration and evading authentication.