Researchers at SenseTime Technology have developed a benchmark for testing facial spoofing detectors.

2020-01-18

Face swapping is a form of deepfake that involves extracting the faces of individuals from existing media and replacing them with the features of other people, typically using artificial intelligence and machine learning. It has been popularized by applications such as MixBooth and SnapChat, and while the underlying technology has made complex image editing more accessible, it has also raised concerns about potential misuse or abuse. Different organizations have compiled manipulated media to support the development of face swapping detection methods, but so far, the number of published samples has been relatively small or too artificial. To address this, researchers from SenseTime Research collaborated with Nanyang Technological University in Singapore to design a new large-scale benchmark for face swapping detection called DeeperForensics-1.0. According to the researchers, all the videos in DeeperForensics-1.0 have been carefully selected, ensuring good quality and diversity. On the surface, they appear more realistic and closer to real-world detection scenarios compared to other datasets. They include different videos or images with compression, blurring, and transmission artifacts, as well as deepfakes discovered in the wild. To construct DeeperForensics-1.0, researchers collected facial data from 100 paid actors of both genders from 26 different countries, ranging in age from 20 to 45. They were asked to turn their heads naturally under nine lighting conditions and speak with more than 53 different expressions. Using an artificial intelligence framework, they ran these data with 1,000 YouTube videos as target videos, swapping the faces of the 100 actors onto 10 targets. They intentionally distorted each video in 35 different ways to simulate real-world scenarios, aiming for the best experimental results. To evaluate the quality of DeeperForensics-1.0 compared to other publicly available datasets, the researchers had 100 computer vision experts rank the quality of a subset of the videos. They reported that DeeperForensics-1.0 consistently outperformed FaceForensics++, Celeb-DF, and other popular deepfake detection corpora in terms of realism. In future work, the research team plans to delve deeper into the field of forensic analysis and collaborate with the research community to establish evaluation metrics for face swapping detection methods.

RECENT AI NEWS

RECENT AI TOOLS