They present a deepfake detection system for faces with masks

0
40
destacada 1.jpg
destacada 1.jpg

The last few years have shown us the powerful audiovisual tool that deepfakes are. Being an intervened material that can hardly be detected as such under a daily gaze, they are a dangerous tool in the wrong hands.

For the same reason, different deepfake detection models have also been developed, taking advantage of the benefits of deep learning and artificial intelligence. However, despite the good results obtained, these systems have a great weakness in common: the correct analysis of faces with masks.

AI to detect deepfakes in covered faces

Intervened image detection systems have shown a high level of performance against different samples of deepfakes. However, the health crisis unleashed by the global outbreak of Covid-19 naturalized the use of masks, which by covering an important part of the face, make it difficult to fully analyze a face.

In the current context, considering the increased probability of encountering a photograph or video of a person wearing a face mask, a team of researchers ventured into the search for a solution so that these detection systems work properly. For that, they presented two new approaches to generate training data sets.

In the first, called “face patch,” models are trained with covered areas of the face, which censor the mouth and nose portions of the deepfakes’ faces. In the other alternative, called “face culture,” models are trained with fake images cropped onto the face mask.

The detected need for a new data set that contemplates the use of masks was addressed by combining generated images and real photographs of faces with masks.

According to the researchers in their development report, it was possible to demonstrate that the facial culture method outperforms the facial patch.

SEE ALSO  WhatsApp Web is working on a new function to protect your conversations

The team behind this development also reported that, after numerous experiments, they succeeded in proving that this method can be used to effectively determine false faces with a face mask in the real world.

The emergence of deepfakes, which are synonymous with the generation and manipulation of hyperrealistic facial images, have given rise to numerous social and ethical problems, such as the invasion of privacy, security threats and malicious political actions. These situations prompted the development of detection methods for deepfakes, including for forensic analysis.

Now, with this new proposal, these detection and analysis mechanisms could be more prepared to function successfully today.