Mitigation of imperceptible fingerprints through targeted optimization

Determine whether imperceptible passive fingerprints in speech generators can be mitigated through targeted optimization using dedicated datasets.

Background

The paper presents the Perceptible-Imperceptible Passive-fingerprint Diagnostic Protocol (PIPDP) and finds that imperceptible fingerprints provide persistent attribution cues across multiple speech generators. Because these fingerprints are not consciously perceived and are rarely represented in dedicated supervised datasets, the paper identifies the creation of such datasets and the assessment of targeted optimization as a direction for resolving whether these attribution-relevant traces can be reduced or controlled.