msaidov/audioseal-robust-sgmse-16bits
Updated • 58
AudioSeal Robustness to Diffusion Attacks
AudioSeal generators fine-tuned against diffusion watermark-removal attacks (SGMSE, AudioLDM). Built at PSIML 11 in Serbia.
Note Trained against SGMSE (OU-VE SDE speech enhancement), 50/50 identity/attack, on a 10h LibriSpeech subset. AudioLDM held out as the unseen-attack probe.
msaidov/audioseal-robust-audioldm-16bits
Updated • 27Note Trained against AudioLDM latent-diffusion resynthesis, 50/50 identity/attack, on full LibriSpeech train-clean-100 with gradient normalization. SGMSE held out.