In the paper, HE-Diffusion is introduced as a tailored encryption framework designed to align with stable diffusion architecture. It focuses on protecting the denoising phase of the diffusion process by proposing a novel min-distortion method for efficient partial image encryption. The adoption of sparse tensor representation enhances computational operations' efficiency, leading to successful implementation of HE-based privacy-preserving stable diffusion inference. Experimental results show significant speedup compared to baseline methods, maintaining performance and accuracy on par with plaintext counterparts. The integration of advanced cryptographic techniques with generative models paves the way for privacy-preserving and efficient image generation in critical applications.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Yaojian Chen... kl. arxiv.org 03-12-2024
https://arxiv.org/pdf/2403.05794.pdfDybere Forespørgsler