We propose the Identity Sensitivity Score (ISS), a per-utterance diagnostic that measures how much an audio deepfake detector’s output changes across different speaker-identity contexts, without requiring ground-truth labels at inference time. Across two detectors and two datasets, misclassified utterances show ISS scores tens of times higher than correctly classified ones, indicating that many detector errors are tied to speaker-identity reliance rather than synthesis artifacts alone.
@article{dar2026probing,title={Probing Speaker Identity Sensitivity in Audio Deepfake Detectors},author={Dar, Daniyal Kabir and Ross, Arun},journal={IEEE/IAPR International Joint Conference on Biometrics (IJCB)},year={2026},archiveprefix={arXiv},primaryclass={cs.SD},}
ICASSP
Impact of Phonetics on Speaker Identity in Adversarial Voice Attack
Daniyal Kabir Dar, Qiben Yan, Li Xiao, and 1 more author
In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026
We analyze adversarial audio perturbations at the phonetic level, showing that they exploit systematic confusions such as vowel centralization and consonant substitutions. These distortions degrade phonetic cues that are critical for speaker verification, leading to measurable identity drift under adversarial attack.
@inproceedings{dar2026impact,title={Impact of Phonetics on Speaker Identity in Adversarial Voice Attack},author={Dar, Daniyal Kabir and Yan, Qiben and Xiao, Li and Ross, Arun},booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},pages={13462--13466},year={2026},publisher={IEEE},archiveprefix={arXiv},primaryclass={cs.SD},}