[Keynote Speaker] Detecting the Undetectable : Robust Defense Strategies Against Audio Deepfakes

講演者：Junichi Yamagishi
会議名：1M-Deepfakes Detection Challenge – ACM Multimedia 2025
主催者：
開催地： Dublin Royal Convention Centre, Ireland
開催日：2025年10月29日
URL：https://deepfakes1m.github.io/2025/program

Recent developments in speech synthesis—particularly voice cloning technologies capable of accurately reproducing speaker identity—have unlocked new opportunities in domains such as entertainment and human-computer interaction. However, these same capabilities present serious security concerns when misused, especially in the context of speaker verification and other biometric authentication systems. In this presentation, we introduce our research on countermeasures against spoofing attacks involving deepfake audio. We begin by discussing ASVspoof, a large-scale speech corpus widely adopted for training deepfake detection models, and present our analysis of detection performance under challenging conditions such as telephony and compressed audio transmission. Given the rapid and continuous evolution of media generation techniques, we also introduce our recent methods for detecting previously unseen synthesis approaches. This includes leveraging a large-scale speech foundation model, Anti-Deepfake, to enhance detector generalization and developing automatic data selection algorithms to dynamically expand deepfake training datasets, thereby contributing to an automatic framework for adaptive model optimization.