Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Research
arXiv CS.AI·1d ago·5 min read
Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection
TL;DR
A new framework for audio deepfake detection combines Adaptive Audio Feature Integration and Mel-Band Adversarial Perturbation, achieving a state-of-the-art Equal Error Rate of 1.51% while reducing model size by 91.6%. This approach enhances detection across diverse spoofing attacks and languages.
✦ Why It Matters
Implement this framework to improve audio deepfake detection in your applications today.
Key Takeaways
- Explore the Adaptive Audio Feature Integration for flexible representation.
- Utilize Mel-Band Adversarial Perturbation for robust learning.
- Consider the Whisper-small model for efficient deepfake detection.
- Test the framework's performance across different languages.
How It Works
The framework employs an Adaptive Audio Feature Integration (AAFI) module that dynamically selects the best latent dimensions from a compression pool, allowing it to adapt to various spoofing attacks. Additionally, the Mel-Band Adversarial Perturbation (Mel-BAP) strategy applies targeted regularization in the mel-spectrogram domain, which helps the model learn decision boundaries that are invariant to perturbations, enhancing its robustness.