AI-röstkloning, är de en människa eller maskin på andra sidan? : En studie om AI-röstkloning och dess förmåga att imitera verkliga röster

Sammanfattning

Recent developments in artificial intelligence have paved the way for a wide variety of tools based on this technology. Among these is AI-based voice cloning. Multiple organizations have developed AI-models with this use in mind, and the development continues at a rapid pace. This, as with many other new technologies, creates new challenges for society. This study examines this dilemma from a cybersecurity perspective, where participants’ ability to identify AI-generated voice material, is tested through a survey, with the goal of gaining quantitative data. The study also included interviews aimed at understanding the participants’ motivations behind their assessments. Finally, an experimental method was used in which different voice clone models were tested against a simulated voice-based biometric authentication system. Results from the study show that AI-generated recordings in most cases were correctly identified by participants, however a significant proportion of incorrect identifications indicates risks of only using human perception as a method of verification. Some variation was observed between the age groups. However, since the groups were unevenly distributed and several age groups consisted of fewer participants than desired, the results should be interpreted with caution and primarily as descriptive patterns rather than generalizable age-related effects. The study also shows differences among the AI models in how difficult it was for the participants to correctly differentiate between the AI-generated voice/audio material. The results indicated no strong correlation between the survey participants’ self-assessed confidence and their actual correctness. The study also shows that language selection could be relevant to the performance of the different models. The results show that the generated Swedish audio/voice material was easier to correctly identify by the participants than the English audio/voice material. The experimental method showed that none of these tested AI-models could bypass the simulated voice-based biometric system, however, differences in performance were recorded In summary, the study shows that AI-based voice cloning should be considered a potential cybersecurity threat. This applies both from a social engineering perspective and as a possible future attack vector against voice-based systems.