ConceptModel Inversion
Model Inversion Attack
A privacy attack method in which an attacker repeatedly queries a machine learning model, such as a TTS endpoint, and observes its outputs to infer sensitive information in the training data, such as a speaker's voice characteristics or fragments of original recordings.
Timeline (last 90 days)
Oct 17
A model inversion attack is a privacy attack in which an attacker repeatedly queries a TTS endpoint and uses the model's outputs to infer sensitive information in the training data; defenses include limiting query frequency, adding noise to outputs, implementing query auditing, and avoiding retaining voice samples of identifiable individuals in training data.
Unverified50%