Are deepfakes and voice cloning detectable?


When technologies evolve, so does cybercrime. Fraudulent emails are no longer enough, so new and more sophisticated techniques are being implemented to blackmail or embezzle sums of up to several million euros.

To begin with, let’s remember what deepfakes and voice cloning are. Deepfakes are photo or video montages where a person’s face is replaced by another. In the case of a video, it can be associated with voice cloning, that is to say the reproduction of a voice using artificial intelligence to make it say something else. From there, it’s easy to guess how these technologies can be used by malicious people.

Different objectives, different techniques

Currently, scammers using this type of technology favor audio cloning, which is more difficult to detect. In 2020, a bank manager authorized transfers to a person he thought was the CEO of a company he knew. In truth, it was a crook who had cloned his voice. The result: $ 35 million was embezzled.

Video deepfake is used more for blackmail and intimidation, especially against women. One of the best-known cases is that of Indian freelance journalist Rana Ayyub. In 2018, because she had denounced the impunity of pedophiles in her country, a pornographic deepfake accompanied by her phone number toured the country. The United Nations had to intervene with the Indian government to end the harassment.

Sophisticated, but flawed scams

Video deepfakes are relatively new, and while they keep getting better, it’s still possible to see the deception. The Spot the Deepfake site allows you to check your knowledge and adopt new reflexes. It shows that the eyes can be blurry or blink oddly for example. Glasses can also be complex to reproduce for deepfake software, which creates artifacts (irregularities). It is also possible to analyze videos and photos thanks to extensions like WeVerify and InVID, intended for journalists to do fact-checking. They allow you to see if the photo or video has already been published elsewhere, to zoom in or to check the metadata.

A deepfake has been created from comedian Tom Cruise’s TikTok account. © DR

As audio cloning is more complicated to detect without specialized software, the best solution remains critical thinking. Take into account the context, intentions and potential consequences of the message to know if it is false. For example, if a person or business asks for a large sum of money insisting on the urgency of the situation, it is probably a scam. It is sufficient to contact this person by another means to confirm this request or, if necessary, to warn him of the theft of his identity.

Media, institutions, companies: the response is getting organized

WeVerify and InVID are good examples of cooperation against deepfakes. Indeed, they were created by AFP’s MediaLab with funding from the European Union’s Horizon 2020 research and innovation program. Across the Atlantic, Facebook, Microsoft and several universities organized the Deepfake Detection Challenge in 2019. As the name suggests, they posted 115,000 videos for participants to review in order to come up with computer models to detect deepfakes. To give a more realistic experience, the quality of the videos and special effects was very variable. For this first competition, the best model was right 65% of the time. There is therefore room for improvement.

In France, the fight against deepfakes takes place in a particular context: that of the presidential election of 2022, threatened with foreign interference through disinformation campaigns on social networks. For this, the government founded in October 2021 an agency for vigilance and protection against foreign digital interference, Viginum.

Leave A Reply

Your email address will not be published.