EAB event goes deep on technical details of voice deepfake detection

The latest online workshop from the European Association for Biometrics (EAB) covers the proliferation of voice deepfakes and adversarial attacks, which exploit vulnerabilities in machine learning to pose a significant risk to biometric security.

The two-hour event features presentations from the National Institute of Informatics (NII) in Japan, and France-based organizations Whispeak and EURECOM. Topics covered include advances in voice deepfake generation and detection; convolutional adversarial attacks; and the challenges of voice deepfake detection in real-world scenarios.

Xin Wang offers a brief history of voice generation and its evolution over the last 40-50 years, followed by a highly technical breakdown of machine learning techniques employed to train large language models (LLM). His key takeaways are that voice generation techniques are improving, and presenting more risks of misuse. In terms of voice deepfake detection, binary classification is still the dominant model, and generalization remains a challenge.

French voice biometric authentication provider Whispeak offers an examination of deepfake detection in the wild, so to speak. Presenter Tony Marteau, an AI research scientist for Whispeak, says “deepfake detectors are commonly trained and evaluated on clean benchmark-style conditions and tested with similar data sets.”

That doesn’t line up with what happens in real-world deployments. Detectors have trouble generalizing acoustic variation and other major parameters that influence detection. Marteau says that diversification of training data is key in achieving better generalization.

Adversarial attacks get a thorough technical breakdown from EURECOM’s Michele Panariello, who computes his way from Malafide attacks through the Malacopula training process and beyond.

Article Topics

biometrics  |  biometrics research  |  deepfake detection  |  deepfakes  |  EAB  |  EAB 2025  |  European Association for Biometrics  |  voice authentication  |  voice biometrics  |  Whispeak