Deepfakes are a growing concern in forensic image and video analysis. AI-based image generation and tampering technologies are becoming more advanced and widely accessible. Amped Software has been dedicating significant research and resources to provide users of Amped Authenticate with effective tools to combat this emerging threat.
Amped Software first addressed this challenge with the release of the Face GAN Deepfake filter, which was developed to detect fake facial images generated by systems like thispersondoesnotexist.com. However, recognizing that this filter was limited to GAN-generated faces, Amped is now taking a significant step forward with the launch of the Diffusion Model Deepfake filter.
Diffusion Model Deepfake Filter
Diffusion models lie at the heart of the most modern AI-based synthetic image generation systems, such as Midjourney, DALL-E, and Stable Diffusion. These systems, commonly referred to as “text-to-image”, create images based on textual descriptions/prompts. For instance, an image of a car on a sunny day can be generated from just a few words. While these technologies are impressive, they present a significant challenge in the field of digital forensics.
Amped Software has closely monitored the developments in this field and identified the research by Cozzolino et al. as a highly promising approach for detecting AI-generated images. The method uses CLIP (Contrastive Language-Image Pre-training) features extracted from the image, and a Support Vector Machine (SVM) classifier to assign the image to one of several classes. With the new Diffusion Model Deepfake filter, this advanced technology is now integrated into Amped Authenticate.
Amped Software trained the system using a large dataset that includes images from common text-to-image generators like Midjourney, Stable Diffusion, and DALL-E, as well as real photographs. When users process an image with the Diffusion Model Deepfake filter, they will receive a tabular output displaying the confidence score assigned to each class. For user convenience, the class with the highest score is highlighted in the “Predicted Class” row.
Additionally, when the predicted class is one of the diffusion models, the Diffusion Model Deepfake filter will alert users by turning red on the filter panel.
Key Considerations for Users
While the Diffusion Model Deepfake filter offers a powerful tool for detecting synthetic images, Amped Software emphasizes that no solution is perfect. Even though the classifier has been trained on a wide range of data, images can be misclassified. A high confidence score does not always guarantee correctness, and the system can sometimes be wrong. Nevertheless, internal testing and experimental validation, both in Amped and in the original paper, have confirmed the method’s strong performance.
Amped Software recognizes the rapid pace at which commercial text-to-image services evolve. Therefore, the software company has implemented a process to ensure the Diffusion Model Deepfake filter is continuously updated in future releases of Amped Authenticate. Amped encourages users to keep their SMS service running to receive these crucial updates.
A Word on AI Use
Amped Software remains committed to promoting user awareness and responsible AI use. Amped Authenticate includes several model-based, fully explainable methods. However, Amped acknowledges that AI-based detectors currently deliver the best performance for detecting deepfakes. Indeed, Amped’s position is that AI can be used for image and video analysis, provided some safeguards are in place.
However, Amped Software cautions against relying solely on AI results. Suppose the Diffusion Model Deepfake filter flags an image as a potential deepfake. In that case, users are encouraged to deepen their analysis using other tools within Amped Authenticate, such as checking shadows, reflections, and Fourier spectrum. Combining several analyses is the best approach to strengthen conclusions.
Conclusion
The introduction of the Diffusion Model Deepfake filter significantly enhances Amped Authenticate’s capabilities in identifying deepfakes. Amped Software acknowledges that this is only the beginning, as diffusion models will continue to evolve. To stay ahead of these advancements, Amped is committed to ongoing research and updates to improve the filter in future releases.
Amped Software invites users to subscribe to its blog and follow its social media channels to stay informed about the latest developments in image and video forensic analysis. To learn more about detecting deepfakes with the Diffusion Model Deepfake Filter, read this article on Amped’s blog.
