A Survey of Model Inversion Attacks on Image Domain

초록

The widespread deployment of deep learning models in sensitive applications has raised increasing concerns about potential privacy risks. Among them, the Model Inversion Attack (MIA) stands out as a notable threat, aiming to reconstruct data samples representative of the model's private training data. These risks are particularly concerning in the image domain, where successful attacks may lead to the recovery of recognizable faces or sensitive medical images. In response to the growing research interest in this area, this paper presents a comprehensive survey of MIAs applied to image data. We propose a systematic taxonomy that categorizes attacks by threat models, technical attributes, and core methodologies. Additionally, we discuss widely used evaluation metrics and outline the landscape of existing defenses against MIAs.

키워드