Protecting Emotional Privacy in Voice Data
Using Simple Audio Editing as Defense Against LLM Emotion Detection
This research introduces user-friendly privacy protection through familiar audio editing techniques that defend against AI systems attempting to detect emotions in voice data.
- Audio manipulations like pitch shifting and time stretching significantly reduce emotion detection accuracy
- These defenses are accessible to users without technical expertise
- Protection maintains voice intelligibility while obscuring emotional content
- Defenses work against various modern LLM attack models, including custom-trained emotion detection systems
As voice technologies become ubiquitous, these practical defenses help users maintain emotional privacy without sacrificing usability - a critical balance for wide adoption of privacy measures across voice applications.