Generalizable speech deepfake detection via meta-learned LoRA

2025-02-15

AI Summary

This work addresses the poor generalization of deepfake speech detection models to unseen attack types. We propose a meta-learning framework specifically designed for cross-attack generalization. To our knowledge, this is the first approach that integrates Model-Agnostic Meta-Learning (MAML) with Low-Rank Adaptation (LoRA) to model shared representational structures across diverse synthetic speech sources, enabling zero-shot transfer detection without access to target-attack samples. Our method jointly optimizes speech representation fine-tuning and multi-source forgery distribution modeling, substantially improving robustness against previously unseen generators—including those trained with different random seeds. On standard benchmarks, it achieves a 12.7% average accuracy gain over unknown attacks, outperforming conventional fine-tuning and ensemble baselines. The framework establishes a novel, transferable, and lightweight generalization paradigm for deepfake detection.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation
• Computer Vision: Generative Adversarial Networks (GANs) for Vision

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Generalizable deepfake detection can be formulated as a detection problem where labels (bonafide and fake) are fixed but distributional drift affects the deepfake set. We can always train our detector with one-selected attacks and bonafide data, but an attacker can generate new attacks by just retraining his generator with a different seed. One reasonable approach is to simply pool all different attack types available in training time. Our proposed approach is to utilize meta-learning in combination with LoRA adapters to learn the structure in the training data that is common to all attack types.

Problem

• Generalizable deepfake detection
• Meta-learned LoRA adapters
• Handling distributional drift in attacks

Innovation

• Meta-learning enhances detection
• LoRA adapters improve generalization
• Common structure across attack types