Researchers explore supervised contrastive learning for deepfake audio detection

Researchers have explored supervised contrastive learning techniques to improve deepfake audio detection. Their study focused on varying similarity metrics, such as cosine and angular similarity, and different methods for negative scaling using a cross-batch queue. The findings indicate that cosine similarity with a delayed queue achieved the best performance on specific evaluation datasets, while angular similarity showed promise with reduced reliance on large negative sample sets. AI

IMPACT Offers improved methods for detecting synthetic audio, potentially enhancing security and trust in audio-based systems.

RANK_REASON Academic paper detailing a controlled study on supervised contrastive learning for deepfake audio detection.

Read on arXiv cs.LG →

• ASVspoof 2019 LA
• ASVspoof 2021 DF/LA
• Deepfake audio detection
• Supervised contrastive learning
• wav2vec2 XLS-R

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →