Digital Image Forensics: A quantitative & qualitative comparison between State-of-the-art-AI and Traditional Techniques for detection and localization of image manipulations

Code

• github.com/UHstudent/digital_image_forensics_thesis
  none★ 3

Abstract

With the rise of realistic AI-generated images and continuously advancing photo-editing software, it has become increasingly difficult to reliably distinguish between authentic and manipulated images. Using Digital Image Forensics, the primary objective of this work is to conduct a comprehensive quantitative and qualitative study that compares traditional forensic techniques to a state-of-the-art AI based approach. A practical underpinning for this work is the development of a toolset capable of providing reliable and transparent evidence regarding the authenticity of an image. Results indicate that the developed algorithms were implemented correctly. The quantitative comparison suggests that the combined traditional techniques outperform a recent state-of-the-art AI network by an average of 17% more detections in scenarios where 0% False Positives are allowed. However, the study acknowledges potential biases in the validation process and further experimentation is necessary to ascertain the reliability of these findings. A qualitative comparison suggests that traditional techniques are more dependable than AI. This work emphasizes the importance of developing reliable digital image forensic tools and outlines a future vision where AI can be utilized in a key support role to assist forensic analysts. Guided by an open-source philosophy, each algorithm was successfully integrated into Sherloq, an established open source image forensic toolset, which has garnered positive feedback from the community.

Tasks

Reproductions