FSD 2.0: Improved Fake Song Dataset for Unknown-domain Deepfake Detection

With the enhancement of song synthesis technology, singing deepfake detection have emerged as critical research domains. Current fake song detection methods demonstrate promising performance in controlled environments, with high accuracy in seen-domain. However, these methods do not focus on unknown domains, resulting in poor adaptability to new conditions. To address this limitation, we introduce FSD 2.0, a novel fake song dataset that surpasses existing resources in scale, diversity, and methodological coverage. Our experiments indicate that models trained on FSD 2.0 achieve 48% improvement in unknown domain detection scenarios, compared with trained solely on the original FSD dataset. These results advance the field’s capability to identify synthetic singing voices across diverse and previously unseen contexts.