Traceback Translators Against Forgetting in Continual Fake Speech Detection

2026-07-14

AI Summary

This work addresses the catastrophic forgetting problem faced by fake speech detectors in continual learning scenarios, where model updates degrade performance on previously learned data. To mitigate this issue, the authors propose an anti-forgetting framework that combines a frozen detector with a domain translation network. By mapping features of newly encountered data back to the original training domain, the method preserves high detection accuracy for both historical and newly added fake speech samples without requiring retraining of the backbone model. Compared to conventional retraining strategies, the proposed approach substantially reduces computational overhead while consistently maintaining robust detection performance across multiple rounds of continual learning, demonstrating superior efficiency and stability.

Abstract

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 nowadays widely considered feasible strategies for updating models to new datasets, but they also 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 the achievement of high detection rates with respect to traditional retraining, while minimizing the computational effort and preserving the detection accuracy on previous data.

Problem

• continual learning
• fake speech detection
• catastrophic forgetting
• generative models
• domain translation

Innovation

• continual learning
• catastrophic forgetting
• domain translation
• fake speech detection
• traceback translator