Using POWER Technology to Detect Deepfakes
Published on Tuesday 21 January 2020
Ganesan Narayanasamy, OpenPOWER Leader in Education and Research, IBM Systems
Our society is struggling with deepfake images and videos and the harmful impact that they can have by spreading misinformation. Detecting these malicious efforts will become even more difficult as the technology becomes more advanced. As deep-fake videos continue to influence public opinion, it’s becoming increasingly important to develop technology that can detect and reveal deepfakes as the false information they are.
This is what makes the work of Pranjal Ranjan, Sarvesh Patil, Badhrinarayan Malolan, Ankit Parekh & Saksham Singh at Veermata Jijabai Technological Institute (VJTI) Mumbai so significant and exciting. The students presented their work on deepfake detection at the 26th IEEE International Conference on High Performance Computing, Data, And Analytics held in Hyderabad, India last month. The Conference serves as a platform for showcasing current work by researchers in the field of high power computing.
To build a deepfake detection program, the students used a number of videos involving facial reenactments, where the facial movements and words from one person are swapped with those of another, creating a video where a person appears to be saying something that they did not, in fact, say.
The students’ video, at the bottom of this article, for example, demonstrates how deepfake technology can be used to edit a video of former US President Barack Obama as a public address, so that the President appears to be saying something originally said by actor and director Jordan Peele.
By examining the technology behind facial reenactment, the students were able to detect the exact location of the facial manipulation in a fake image or video, and therefore reveal the deepfake.
The students partnered with the University of Oregon to receive access to their Power9 systems to train their models. By using the computational power of the IBM AC922 server, containing 4 NVIDIA Tesla V1000 GPUs, the students found a 30% performance boost over other similar setups. This allowed them to train their model more efficiently and quickly.
Learn more about the students’ work and demonstration here: