Face images created artificial intelligence (AI) are so realistic that even “super recognizers”—an elite group with exceptionally strong facial processing abilities—are no better than chance at detecting fake faces.
People with typical recognition abilities are worse than chance: they are more likely to think that AI-generated faces are real.
“I think what was encouraging was that our fairly short training procedure significantly improved performance in both groups,” says the study’s lead author. Katie GrayAssociate Professor of Psychology at the University of Reading in the UK, told Live Science.
Surprisingly, training improved accuracy by the same amount in super-recognizers and typical recognizers, Gray said. Since super recognizers are better at initially recognizing fake faces, this suggests that they rely on a different set of cues rather than just rendering errors to identify fake faces.
Gray hopes that in the future, scientists will be able to use the improved detection skills of super-recognizers to better recognize images generated by AI.
“To best detect synthetic faces, it may be possible to use AI detection algorithms with a human-in-the-loop approach, where that human is a trained SR. [super recognizer]”, the authors write in the study.
Deepfake detection
In recent years, the Internet has seen an influx of AI-generated images. Deepfake faces are created using a two-stage artificial intelligence algorithm called generative adversarial networks. First, a fake image is created based on real images, and then the resulting image is carefully checked by a discriminator that determines whether it is real or fake. Through iteration, the fake images become realistic enough to get past the discriminator.
These algorithms have now improved to the point that people are often fooled into thinking that fake faces are more “real” than real faces—a phenomenon known as “hyperrealism“
As a result, researchers are now trying to develop training programs that can improve people's ability to detect AI faces. These trainings indicate common rendering errors in AI-generated faces, such as those with a middle tooth, strange hairline, or unnatural skin texture. They also emphasize that fake faces tend to more proportional than real ones.
In theory, so-called super-recognizers should be better at recognizing counterfeits than the average person. These super recognizers are people who excel at face perception and recognition tasks, in which they might be shown two photographs of strangers and asked to determine whether they are the same person or not. But to date, few studies have examined super recognizers' ability to detect fake faces and whether training can improve their performance.
To fill this gap, Gray and her team conducted a series of online experiments comparing the performance of a group of super recognizers with the performance of typical recognizers. Super Contributors were recruited from Greenwich Facial and Voice Recognition Laboratory volunteer database; they were in the top 2% of participants in tasks where they were shown unfamiliar faces and had to remember them.
In the first experiment, an image of a face appeared on the screen, which was either real or computer-generated. Participants had 10 seconds to decide whether the face was real or not. The super recognizers performed no better than if they guessed randomly, recognizing only 41% of the AI faces. Typical recognizers correctly identified only about 30% of fakes.
Each cohort also differed in how often they judged real faces to be fake. This occurred in 39% of cases for super recognizers and approximately 46% for typical recognizers.
The next experiment was identical, but included a new group of participants who underwent a five-minute training session in which they were shown examples of errors in AI-generated faces. They were then tested on 10 faces and provided real-time feedback on the accuracy of fake detection. The final phase of training involved reviewing rendering errors to look out for. Participants then repeated the original task from the first experiment.
Training significantly improved detection accuracy: super recognizers recognized 64% of fake faces, while typical recognizers recognized 51%. The proportion of cases in which each group incorrectly labeled real faces as fake was about the same as in the first experiment: super recognizers and typical recognizers judged real faces as “fake” 37% and 49% of the time, respectively.
Trained participants tended to take longer to study the images than untrained participants, with typical recognizers slowing down by about 1.9 seconds and super recognizers slowing down by 1.2 seconds. Gray said this is a key message for anyone trying to determine whether a face they see is real or fake: Slow down and carefully examine the facial features.
However, it's worth noting that the test was conducted immediately after participants completed the training, so it's unclear how long the effects will last.
“The training cannot be considered a long-term and effective intervention because it has not been retested,” Maike Ramona professor of applied data science and an expert in facial processing at the Bern University of Applied Sciences in Switzerland, wrote in a review of the study conducted before it was published.
And because the two experiments used different participants, we can't be sure how much the training will improve human detection skills, Ramon added. This would require testing the same group of people twice, before and after training.
