Excited to share our new paper in INTERSPEECH '24 on embedding signatures in audio to detect and prevent #deepfake audio. Read our paper here: arxiv.org/abs/2407.00913.
1
1
4
506
Imagine a hacker fakes your voice, calls your bank, or impersonates you on social media. Our unique audio signatures secure your voice, making it easy to detect and prevent #DeepFake attacks. Your voice interactions remain secure and authentic.
1
70
We analyzed DeepFake audio spectra using CommonVoice, LibriSpeech, and VoxCeleb datasets. By converting audio into spectrograms and segmenting into bins, we observed significant high-frequency attenuation in the clones compared to the original signals, as shown in the figure.
Jan 7, 2025 · 12:47 AM UTC
1
60
The more orthogonal data (signatures), the better protection against clones. Our method outperforms existing methods by improving verification accuracy by 81% and exceeding recent works by 71% and 42%. Adding DP noise secures signatures with only a 4% accuracy drop.
1
24
DP noise prevents adversaries from reverse-engineering signatures. Even if an adversary accesses multiple signed audio, they cannot reconstruct the original signature, ensuring robust protection against sophisticated cloning attacks.
1
1
258
Check out our code here: github.com/UTAustin-SwarmLab…. @bittensorai subnet anyone?
160