Deepfakes detect Zoom-bombing culprits
In 2020, Zoom became the video conferencing platform of choice for many people forced to work from home and socially distance because of COVID-19.
With this popularity came a rise in trolling and ‘Zoom-bombing’, unwanted disruptions and the hijacking of many meetings and virtual events worldwide.
In a collaboration between Monash University and the Indian Institute of Technology Ropar, researchers developed FakeBuster - a deepfake detection tool which enables a user to uncover if another person’s video is being manipulated or spoofed during a video conference. Developed independently of video conferencing solutions, this deep learning-based tool has been tested using both Zoom and Skype.
With the ongoing rise in popularity of video calls, FakeBuster offers a solution which detects possible intruders in real time, allowing meetings to run without interruption.
Project Lead Dr Abhinav Dhall from the Faculty of Information Technology (IT) at Monash University, says the sophisticated nature of Artificial Intelligence (AI) has led to manipulated media content being readily available.
“Manipulated media content is becoming increasingly difficult to recognise and is making its way into video calls as fake news, facial manipulation and pornography. This is having major repercussions across a number of video conferencing platforms,” Dr Dhall said.
“In an environment where video calling is the quintessential mode of personal and professional communication, FakeBuster allows participants to monitor and verify the authenticity of other participants as it happens.”
The tool uses a 3D convolutional neural network for predicting video segment-wise fakeness scores. The network is trained on a combination of datasets such as Deeperforensics, DFDC, VoxCeleb and deepfake videos created using locally-captured images from video conferencing scenarios. This leads to different environments and changes in the dataset, which improves the generalisation of the deepfake network.
To check for imposters, the tool scans for faces on the computer screen and the user selects a face from a video feed displayed via the video conferencing tool. Deep learning-based deepfake detection then generates a fakeness score for a one second-long segment of the selected person. Ultimately the results are collated over the segments and presented to the user, helping to determine if a person’s video feed is a deepfake.
The tool can be used in a number of scenarios including online proctored exams, online job interviews and verification of videos on social media. When there are multiple users in a video conference, an imposter can easily go unnoticed.
The team of researchers are working on improving the tool so that it works better against different deepfake generation techniques.