Unveiling DeepFakes: An MIT Initiative to Counteract AI-manipulated Videos

Deep diving into the world of AI-manipulated videos, the MIT Media Lab's 'Detect Fakes' project presents a comprehensive guide to understanding and identifying DeepFakes.

MIT Media Lab created the Detect Fakes project to educate the public about DeepFake videos and how to identify them. DeepFakes are artificially manipulated videos that can be highly realistic, making it challenging for ordinary individuals to distinguish them from unaltered footage. The Detect Fakes initiative aims to identify strategies to combat this AI-enabled misinformation and foster critical thinking about media consumption.

The believability of DeepFakes was tested in a competition on Kaggle known as the Deepfake Detection Challenge (DFDC). This challenge, sponsored by AWS, Facebook, Microsoft, the Partnership on AI’s Media Integrity Steering Committee, and academics, encouraged researchers worldwide to develop innovative techniques for detecting deep fakes and manipulated media. The winners of this competition were awarded a prize of $1,000,000.

Rather than focusing on refining the best AI model for the Kaggle competition, the Detect Fakes project aimed to explore strategies and techniques for raising public awareness of DeepFake technology. The project's creators hypothesized that exposing the characteristics of DeepFakes and offering people the experience of identifying subtle algorithmic manipulations could enhance their ability to spot a wide range of video manipulations in the future.

To this end, the Detect Fakes website was set up to showcase thousands of carefully selected high-quality DeepFake and real videos from the DFDC dataset. The website also includes 32 videos from the Presidential Deepfakes Dataset.

The Detect Fakes experiment provides an opportunity to gain a deeper understanding of DeepFakes and evaluate one's ability to distinguish between real and fake videos. While there isn't a single surefire method for identifying a DeepFake, several common artifacts can be watched out for. These include:

1. Facial transformations, such as high-quality DeepFakes, usually alter the face.

2. Skin texture inconsistencies include smoothness, wrinkles, and age-related features.

3. Shadows around the eyes and eyebrows may not adhere to natural physics in DeepFakes.

4. Glare on glasses, which may not change appropriately with movement due to inaccurate lighting physics in DeepFakes.

5. Facial hair, which may not appear natural in DeepFakes.

6. Facial moles that may not look real.

7. Blinking patterns that may be abnormal.

8. Lip movements, as some DeepFakes, are based on lip-syncing and may not appear natural.

These eight points are designed to guide users in identifying DeepFakes. While high-quality DeepFakes are not easy to recognize, practice can help build intuition for determining what's real and fake.

Source: https://www.media.mit.edu/projects/detect-fakes/overview/