Google adds massive image database to fight against deepfakes
Google has partnered with Jigsaw to produce and deliver a massive database of visual deepfakes that is now part of the FaceForensics benchmark created by the Technical University of Munich and the University Federico II of Naples, the tech giant announced in a blog post.
First detected in 2017, deepfakes are generated by deep generative models that alter video and audio. There are countless generation methods in the wild, including open-source, which could become a threat at any given time if in the wrong hands. Deepfake attacks have claimed their first victim after a fake voice recording was used to impersonate the executive of a British energy company, and manipulate an employee to make a wire transfer worth nearly a quarter-million dollars.
The database comprises of hundreds of recorded videos which were manipulated with widely available deepfake generation methods to create thousands of deepfakes.
“Detecting deepfakes is one of the most important challenges ahead of us,” tweeted Google CEO Sundar Pichai. “Following our release of a synthetic audio dataset in Jan, we’re releasing a large dataset of visual deepfakes to support researchers working on synthetic video detection.”
The data can be downloaded for free from the FaceForensics github page.
In a recent interview with the Financial Times, Pichai warned against reactionary or overly broad regulation of artificial intelligence, saying that existing rules may be sufficient. He suggested that “smart regulation” balancing the protection of citizens and innovation is the right approach for AI. Pichai’s goal is to avoid repeating tech industry mistakes of the past by working in partnership with governments.
With the 2020 U.S. presidential election approaching, concern over deepfakes for political manipulation and their impact on the public has grown.
The threat of deepfakes to social discourse has motivated a group of stakeholders including Facebook, the Partnership on AI, Microsoft, and academics from the U.S. and UK to organize a Deepfake Detection Challenge (DFDC), to introduce a new data set to develop technology that identifies video altered with artificial intelligence.