Carding
Professional
- Messages
- 2,871
- Reaction score
- 2,467
- Points
- 113
Scientists from London have shown how difficult it is for people to recognize artificial speech.
Deepfakes are synthetic media that mimic the voice or appearance of a real person. They can be used for manipulation and deception and are one of the main threats associated with the development of AI.
Scientists from London tested how well people can recognize deepfakes. They asked 529 people to listen to real and fake audio recordings in English and Chinese created using the speech synthesis algorithm (TTS). The researchers generated 50 samples of deepfake speech in each language that did not match the samples used to train the algorithm. Listeners had to determine which ones were artificial.
The results showed that people correctly identified forgeries only 73% of the time, and language did not affect accuracy. Even after students were given examples of deepfakes to raise awareness, their results didn't improve much.
Scientists have concluded that people cannot reliably identify deepfake speech, even if they have received training that was supposed to help them distinguish artificial content from real content.
The researchers also noted that deepfakes will become more realistic and difficult to recognize as speech synthesis algorithms evolve. For example, if earlier thousands of samples of a person's voice were needed to create a deepfake, now only a three-second fragment of their speech is enough.
The difficulty of identifying deepfake speech highlights their potential for abuse and points to the need to develop protective measures.
A senior researcher at the Citizen Lab at the University of Toronto, wrote that current trends have huge implications for phishing and fraud.
He noted that good deepfakes and creative scammers can combine their efforts and quickly adapt to the situation. The researcher believes that phone fraud and phishing are based on very short feedback loops, in which operators instantly learn from their mistakes and successes.
Deepfakes are synthetic media that mimic the voice or appearance of a real person. They can be used for manipulation and deception and are one of the main threats associated with the development of AI.
Scientists from London tested how well people can recognize deepfakes. They asked 529 people to listen to real and fake audio recordings in English and Chinese created using the speech synthesis algorithm (TTS). The researchers generated 50 samples of deepfake speech in each language that did not match the samples used to train the algorithm. Listeners had to determine which ones were artificial.
The results showed that people correctly identified forgeries only 73% of the time, and language did not affect accuracy. Even after students were given examples of deepfakes to raise awareness, their results didn't improve much.
Scientists have concluded that people cannot reliably identify deepfake speech, even if they have received training that was supposed to help them distinguish artificial content from real content.
The researchers also noted that deepfakes will become more realistic and difficult to recognize as speech synthesis algorithms evolve. For example, if earlier thousands of samples of a person's voice were needed to create a deepfake, now only a three-second fragment of their speech is enough.
The difficulty of identifying deepfake speech highlights their potential for abuse and points to the need to develop protective measures.
A senior researcher at the Citizen Lab at the University of Toronto, wrote that current trends have huge implications for phishing and fraud.
He noted that good deepfakes and creative scammers can combine their efforts and quickly adapt to the situation. The researcher believes that phone fraud and phishing are based on very short feedback loops, in which operators instantly learn from their mistakes and successes.
