Deepfakes a lidé: dokážeme ještě rozlišit pravou řeč od umělé?

deepfake, synthetic voice, deepfake attacks, human factor

Deepfakes are unfortunately slowly but surely making headlines. They are gaining this position largely due to their negative use in successful attacks. These attacks (e.g., financial scams where the attacker impersonates someone else) primarily target our ability to recognize (or rather not recognize) the deepfake voice from the real one. To improve our ability to protect against these attacks, it is first important to understand the human ability to distinguish a deepfake recording from a genuine one and to determine what factors influence this ability. Although some international research has already addressed this topic, it neglects to take into account the way a real attack takes place. In our research, we reflect this inaccuracy and re- alistically simulate a situation in which an attack could take place. We thus differ significantly from existing studies, but unfortunately the findings are not very favorable in terms of the impact on people. This paper presents recent published attacks using deepfakes as well as the findings of our research, which focuses on the human ability to recognize deepfake recordings and how to improve this ability in the future and prevent attacks using voice deepfakes.