Friends of the Nonverbal Communication Blog, this week we present the paper “Fighting Deepfakes Using Body Language Analysis”, by Yasrab, R.; Jiang, W. y Riaz, A. (2021), in which authors explain what deepfakes are and how to recognize them using body language’s tools.
The development of technology has brought us good and bad things at the same time. Now, for example, almost everybody has access to the internet, the largest information network in history; but there are also dangers from which we must defend ourselves.
One of them is fake videos, called “deepfakes” from now on, which are a threat to people’s privacy.
Until now, we knew that images could be digitally manipulated with tools such as Photoshop, and we are already trained to distinguish, almost always, the fake ones from the true ones.
However, fake videos are becoming more and more convincing due to the quick development of what authors call the “deep learning method”, which has spread and popularized deepfakes. Currently, videos can be manipulated in a way that the face of one person is replaced by another, preserving the original facial expressions and actions, achieving great realism.
Due to the amount of data required to achieve a successful deepfake model is enormous, it is very easy to focus on celebrities and world leaders (presidents, vice presidents, etc.), as there are many images and videos of them on multiple public platforms. Therefore, they are the main targets of apparently real deepfakes.
The danger of deepfakes lies in the fact that disinformation spreads as fast as real news on the internet and therefore, they are used to deceive the general public. Therefore, they are a serious problem for national and social security if used with political objectives.
World leaders like Barack Obama and Donald Trump have sparked contentious debates because of this. For example, former President Trump shared Joe Biden’s deepfake gifs on Twitter in the middle of the campaign for the US presidency.
As we have mentioned, it also affects celebrities. Recently, the actor Tom Cruise appeared on the social network TikTok, to later reveal that he never had an account on this network and it was actually a deepfake.
The most worrying thing is that this technology has become very accessible in recent years, where with a simple application for smartphones, such as Zao or FaceApp, very realistic videos and gifs can be created. Even the aforementioned TikTok has introduced filters that can be used to, as was the case with Tom Cruise, damage someone’s credibility or spread misinformation.
Therefore, it is essential to help the audience so they can identify fake videos and protect people from deepfakes.
The research conducted by the authors aimed to create a new deepfake detection method that can address these emerging threats, as well as improve existing methods.
We have already mentioned that most of the targets of these scams are famous people or world leaders. For this reason, the research focuses on the latter, whose public appearances are usually behind a lectern, leaving the upper part of their body exposed. Authors hypothesize that the movements of this part of the body are radically different for each individual, and deep learning networks can use the language of these areas to identify people and expose forgeries.
This upper body area would consist of a few key points: the eyes, the nose, the neck, the shoulders, the elbows and the wrists. Authors suggest that these points could be used to train a deep learning neural network, so that it internalizes the posture and gestures of each person.
The methods that have emerged recently are divided into fake image detection and fake video detection. Artificial intelligence-based video analysis is used, which is relatively new. However, the existing methods to date are actually transitory, because deepfakes are being perfected more and more, correcting their errors and making detection difficult. Therefore, it is necessary to develop new techniques or improve existing ones. The novelty of the technique proposed by authors is that it includes the upper part of the body, and not only the face.
They propose to train software to learn the pose of the target person, their body language, and detect fake videos. For their study, they chose videos of George W. Bush, Barack Obama, Donald Trump and Joe Biden.
The results were promising. The authors’ model, which could be improved in the future, found with 94.39% accuracy which videos were false and which videos were true. This shows that upper body body language is very helpful when it comes about exposing deepfakes.
To improve this research, authors propose collecting more videos, as training the software with a larger data set could improve its performance.
