Bbabo NET

Science & Technology News

Research: deepfake faces are becoming harder to distinguish from real ones

Researchers Sophie Jay Nightingale and Hani Farid published their paper in the Proceedings of the National Academy of Sciences of the United States, which demonstrates that deepfake faces are becoming increasingly difficult to distinguish from real ones.

Study participants were able to classify real and artificial faces with an average accuracy of only 48.2%, close to a 50% chance.

For their work, the researchers used 400 synthetic faces generated by StyleGAN2. These were 200 female and 200 male deepfakes in various age ranges, differing by race (100 blacks, 100 Caucasians, 100 East Asians and 100 South Asians). Scientists used only images with a uniform background and no obvious rendering artifacts.

The neural network extracted a low-dimensional representation of each face from its database to compare it with a database of real faces and get the most similar deepfake.

In the first experiment, 315 participants, one, classified 128 out of 800 faces as real or synthesized. Participants were able to guess with an average accuracy of only 48.2% of deepfakes.

When classifying real faces, there was a significant correlation between gender and outcomes. The mean accuracy was higher for East Asian male faces than for East Asian female faces. It was also higher for white males than for white females. The study did not find such a significant interaction between race, gender, and selection outcomes for deepfake faces.

In the second experiment, 219 new participants classified 128 faces, but with training and serial feedback. The average accuracy improved slightly to 59% (59.3% for the first set of 64 faces and 58.8% for the second set of 64 faces).

A third experiment was designed to see if there was a difference in the perception of deepfakes and real faces. A total of 223 participants rated the reality of 128 faces from the same set on a scale of one to seven (one for very unreliable and seven for trustworthy). For real shots, the average coefficient was only 4.48 compared to 4.82 for synthetic images. Women's faces received a higher coefficient (4.94) compared to men's (4.36). However, black faces were also more trustworthy than those from South Asia.

In 2021, Microsoft and Facebook announced the start of a competition to develop deepfake recognition tools, which will last until March 2022.

Facebook AI and researchers at Michigan State University have unveiled a method for detecting deepfakes based on reverse engineering a fake image and a system for creating it. The developers plan to make the source code open to facilitate further research.

And a group of scientists from the University of California, San Diego has demonstrated that even the most advanced deepfake detection systems can be fooled. To do this, it is enough to inject input data or adversarial examples into each video frame of the deepfake.

Research: deepfake faces are becoming harder to distinguish from real ones