This blog post highlights groundbreaking research on deepfakes presented at WIFS 2024. Discover how analyzing shadows and reflections in AI-generated images can expose manipulation.
At Amped Software, we’ve always stood at the intersection of applied technology and scientific rigor. Our commitment to advancing forensic video analysis is evident in our industry-leading tools and also in our active contributions to the global scientific community.
We’re proud to share that our latest research on detecting deepfakes was presented a few months ago at the prestigious IEEE International Workshop on Information Forensics and Security (WIFS) 2024. This marks another significant milestone in our journey!
Our scientific paper, “Assessing shadows and reflections consistency in AI-generated images”, has been published in the WIFS 2024 proceedings and is available on IEEE Xplore here.
The paper made two contributions, which are described below.
Evaluating the Consistency of Shadows and Reflections in Diffusion Model Deepfakes
We created 150 visually credible deepfake images with three well-known services based on diffusion models: Midjourney, Dall-E, and Fooocus.
The authors then manually analyzed shadows and reflections in all of them, and the results were peer-reviewed. They revealed that, at the time of writing, Dall-E and Fooocus had a tough time creating consistent shadows and reflections, while Midjourney performed better with both. Nevertheless, 16 out of 40 images with shadows and 4 out of 10 images with reflections were found inconsistent for Midjourney., This confirmeding that geometric analysis is a powerful tool even when dealing with high-quality deepfake images.
Defining a Practical Approach to Reflection Analysis
In the paper, we demonstrated that reflection consistency can be addressed similarly to shadow consistency by exploiting the main geometric properties of the linear perspective, as visually suggested in the image below. This has resulted in the recent addition of the Reflection filter to Amped Authenticate.
This contribution underscores Amped’s continued efforts to address the complex challenges of image and video authentication through rigorous, peer-reviewed research. Subscribe to our blog and follow us on social media so you won’t miss any discoveries from our research team!
