FinVolution Group, a leading fintech service provider, concluded the 9th FinVolution Global Data Science Competition on July 23 . As part of the IJCAI 2024 challenge track, a top international AI conference, this years event spotlighted the use of large language models (LLMs) for effective voice deepfake detection.
The rapid advancement of deepfake technology blurs the lines between AI-generated and real voices, raising concerns about personal privacy and financial security. Participants of the contest employed deep learning and adversarial AI techniques to develop models capable of detecting fake voices from the provided datasets. The finals featured various algorithmic models and training approaches, including LLMs and traditional end-to-end recognition technologies.
The winning team achieved an impressive fake voice recognition rate of over 99% in the preliminary round and nearly 80% in the semifinals. Qiang Lyu, FinVolutions algorithm scientist, attributed the disparity to varying complexities in the datasets.
The preliminary rounds datasets primarily consisted of cloned voices generated by end-to-end TTS (text-to-speech) systems, which were easier to identify, Lyu explained. In the semifinals, however, the datasets included the latest LLM-synthesized voices, re-recorded fake voices, and even mixed real and fake voices in over five languages, including English, French, and Spanish. This complexity made recognition more challenging.
Addressing LLM-generated deepfakes with LLMs
