How people perceive audio deepfakes has become the subject of scientific

[tanjimajuha20]

A person is unable to distinguish up to 25-27% of audio deepfakes created using neural networks from real audio recordings of people's voices — this is the conclusion reached by employees of University College London. According to experts, audio recordings created by neural networks can become a dangerous weapon in the hands of criminals. Details about the risks of audio deepfakes and how to protect yourself from them are in the Izvestia article.

New threat
work by specialists from jamaica whatsapp resource University College London (UCL). They conducted an experiment in which 500 subjects took part, divided into two groups - all of them were native English and Chinese speakers.

One group of participants spent a long time learning to recognize audio deepfakes, while the other half of the subjects did not prepare for the experiment at all. The scientists recorded voice samples of several men and women who spoke these languages ​​and used them to train the VITS neural network.

The researchers then prepared 50 short audio recordings generated using the VITS neural network, which they offered to listen to the participants of the experiment. The latter had to try to answer whether the recording was real or synthesized by artificial intelligence. The authors of the experiment loaded the same 50 audio tracks into the LFCC-LCNN neural network algorithm, specially created to detect deepfakes.

As it turned out, both trained participants in the experiment and ordinary people were unable to correctly identify up to 25-27% of deepfakes. Such results were approximately the same for both English-speaking subjects and native Chinese speakers. Meanwhile, the LFCC-LCNN algorithm was able to correctly recognize all audio tracks synthesized by artificial intelligence.

Deepfakes Mechanics
University College London is confident that the results of the scientific work on audio deepfakes indicate the need to create new approaches and tools to recognize content generated by neural networks, as well as to prevent their illegal use. According to UCL Professor Lewis Griffin, due to the high dynamics with which media content created with the help of artificial intelligence is developing, humanity may face new risks.

"It would be wise for governments and businesses to develop strategies and tools to combat the misuse of these technologies without limiting their legitimate uses," said Professor Griffin.

As Sergey Polunin, head of the infrastructure IT solutions protection group at Gazinformservice, tells Izvestia, today audio deepfakes (like others) are made using machine learning and artificial intelligence technologies that are trained on the source material. Moreover, the more source voice material, the more believable the fake will be.

- To produce audio deepfakes, specialized software, algorithms for voice processing and synthesis, as well as computing resources, such as powerful graphics processors and training models, are required, - noted, in turn, IT specialist and CEO of ProControl Stanislav Sidorov.

According to Sergei Polunin, if there is a sufficient amount of source material of the proper quality, modern voice models are trained quite easily. Moreover, there are entire services that allow you to create audio deepfakes literally "turnkey", without going into technical details.

On the darknet, they offer to create such videos with any person for $20,000.
However, to create believable deepfakes, high technical competence in this matter is still necessary. If it is available, audio recordings created by neural networks can recreate the unique characteristics and intonations of a specific person, which can be used for a variety of purposes.