The new algorithm works by generating a "dissonance score" by detecting discrepancies between the audio and video content. Examples include unnatural face or lip movements, or audio lags.
"The machine learning method we've developed is applying a detection technique similar to watching a foreign film with overlaid audio that is not in sync with the lip movements. This disharmony between the audio and the visual leads the viewer to notice that the video isn't quite right, which is what we're mimicking with the machine learning algorithm," said project lead Abhinav Dhall from Monash University's faculty of information technology.
"By producing a 'dissonance score', the algorithm detects if something isn't quite right in a video and then identifies the exact part of a video that has been manipulated. The machine learning algorithm independently learns from these discriminative features, further advancing its ability to detect future deepfakes."
The approach has been tested on more than 18,000 deepfakes. The success rate for distinguishing between genuine and manipulated videos is 91.5%, which the researchers say is better than other advanced deepfake detection methods.
"Deepfakes are becoming an increasingly major concern worldwide. They build upon the problems created by fake news and pose a huge potential threat to democracy. With the upcoming US election, AI-generated images, audio and video are increasingly affecting our ability to separate fact from fiction in the political sphere. A reliable deepfake detection algorithm is needed now more than ever," said Indian Institute of Technology Ropar associate professor Ramanathan Subramanian.