Fake speech detectors are increasingly challenged by the development of new and more accurate generative models. To cope with this problem, continual learning techniques are now widely considered feasible strategies for updating models to new datasets, but they often lead to decreased performance on previously seen samples (catastrophic forgetting).
In this work, we propose a forgetting-resilient solution based on the adoption of domain translators within a frozen detector, which remaps the new feature spaces into the original ones by means of a traceback translator network. Experimental results show that this strategy enables achieving high detection rates compared to traditional retraining while minimizing computational effort and preserving detection accuracy on previous data.
Blogger's Review: This study offers an innovative approach to address catastrophic forgetting in continual learning, particularly in the realm of fake speech detection. By introducing traceback translators, it effectively enhances model adaptability and accuracy, showcasing the potential and challenges of continual learning.