Indian researchers built an AI to detect deepfake videos
What's the story
AI has not only given us extremely capable technologies but also some problems to worry about.
Case in point: deepfakes or highly-manipulated videos that are created with machine learning algorithms and used for harassing people, spreading misinformation.
They are a major problem, but the good news is, a group of Indian-origin researchers has come up with a solution to detect them convincingly.
Here's how.
Solution
AI developed to catch deepfake videos
Often, deepfake videos revolve around superimposing the face of one person, mostly a celebrity/politician, over another one's body.
The algorithms work so well that it becomes really difficult to spot signs of manipulation with the naked eye.
Now, this is why researchers from the University of California, Riverside, have developed a deep neural network that can detect manipulations at the pixel level.
Working
How the system detects manipulations
According to the research, led by University's engineering professor Amit Roy-Chowdhury, the neural network uses various resampling features, long short-term memory (LSTM) cells to separate manipulated regions from the non-manipulated ones in forged imagery.
And, as it seems, it achieves this detection with incredible accuracy by capturing signs of image alteration like JPEG quality loss, upsampling, downsampling, rotation, and shearing.
Quote
Here's what Roy-Chowdhury said about the system
"We trained the system to distinguish between manipulated and non-manipulated images, and now if you give it a new image it's able to provide a probability that the image is manipulated or not," Roy-Chowdhury said, noting that it also localizes the region where manipulation occurred.
Plans
Currently, it works with still images
As of now, the deep neural network developed by Roy-Chowdhury and his team detects manipulations in still images.
They say "if you can understand the characteristics in a still image, in a video, it's basically just putting still images together one after another."
However, according to them, "the more fundamental challenge would be figuring out whether a frame in a video is manipulated."
Information
Even Adobe is working on tool to detect fake videos
Interestingly, this research isn't the only work being conducted to spot fake content. Adobe recently unveiled a program designed to spot photoshopped faces, while DARPA has already developed AI-based forensic tools to spot fake photos and revenge porn videos.