Core Concepts
MetaCloakは、不正なテキストから画像生成を防ぐための効果的で堅牢な手法を提案します。
Abstract
Text-to-image diffusion models allow seamless generation of personalized images from scant reference photos. Existing poisoning-based approaches perturb user images in an imperceptible way to render them "unlearnable" from malicious uses. MetaCloak addresses the limitations of existing methods by proposing a meta-learning framework and transformation-robust perturbation crafting process. It outperforms other baselines in degrading the generative performance under various training settings with and without data transformations. MetaCloak is practical and can be applied to protect images in a black-box manner against online training services like Replicate.
Stats
Extensive experiments on the VG-GFace2 and CelebA-HQ datasets show that MetaCloak outperforms existing approaches.
MetaCloak can successfully fool online training services like Replicate, demonstrating its effectiveness in real-world scenarios.
Quotes
"MetaCloak resolves the limitations of existing works in sub-optimal optimization and fragility to data transformations."
"MetaCloak outperforms other baselines across all metrics, including SDS, IMS-VGG, IMS-CLIP, CLIP-IQAC, and LIQE."