Continual Speaker Identity Unlearning with Minimal Interference
Researchers have developed a new framework called Cumulative ORThogonal Identity Suppression (CORTIS) to address the challenge of continually unlearning speaker identities from zero-shot text-to-speech (ZS-TTS) models. Existing methods fail when unlearning requests are sequential, as they can revive previously unlearned speakers. CORTIS, however, uses Fisher-information-based parameter masking and orthogonal projection to ensure that once a speaker identity is unlearned, it remains forgotten even with subsequent unlearning requests, without needing access to the previously unlearned data. This approach was demonstrated to be effective with the VoiceBox model, outperforming sequential applications of prior methods. AI
IMPACT This research addresses a critical privacy concern in generative audio models, enabling more robust and sequential unlearning of sensitive data.