Russians were told about fraudsters faking the voices of their acquaintances with the help of AI

Fraudsters in their schemes began to fake the voices of an interlocutor familiar to a person using neural networks. Andrey Rybnikov, a lecturer at the Department of Instrumental and Applied Software at RTU MIREA, told RT about this on May 2.

"The model does not copy the recording, but synthesizes it anew, selecting sounds according to statistical patterns. It accurately conveys timbre and intonation, but does not reproduce all the nuances of live speech. Because of this, there are minor glitches: unnatural pauses, shifted accents and fluctuations in intonation," he said.

Rybnikov added that it can be difficult to recognize such a fake, especially if the message is transmitted by voice message without a visual row. According to him, you can notice this if you pay attention to acoustic details: unnatural pauses in speech, monotonous breathing or its complete absence, and atypical formulations.

Neural networks of love: how scammers use AI on dating sites

Attackers play on the involvement of users of dating resources

On April 24, the Roskachestvo press service said that deepfakes generated by artificial intelligence (AI) are almost impossible to distinguish from real people, and the old methods of detecting them no longer work. One of the main difficulties in AI recognition is creating characters using multiple services. The developers generate a face, facial expressions, movements, voice and intonation in different neural networks, and take several photos of people who look similar to each other, whose features "cross."

AI director and entrepreneur Alexey Rykov said on April 5 that generative neural networks are able to reproduce not only a person's appearance, but also his facial expressions, gestures and behavioral features, which turns deepfakes into a manipulation tool. According to him, when creating realistic images and videos, neural networks rely on real data — photos and videos, from which behavioral patterns are extracted.