Different direction adversarial sample for diffusion model.

Abstract

In recent years, breakthroughs in diffusion model technology have significantly lowered the barrier to high-quality image creation. However, the powerful generative capabilities of such technology have also led to unauthorized plagiarism, a serious infringement of the rights of the original creators. This study focuses on the image-to-image scenario of diffusion models, and two contributions are proposed. First, the existing MIST algorithm is simplified. Through theoretical analysis, the perturbation process generated on the basis of semantic loss is approximated by textual loss, thereby unifying the two perturbation methods and facilitating the optimization of the joint loss. Second, a directional iterative enhancement algorithm and RUDDER are proposed in combination, and a simple derivation is performed to establish INV(S). The experimental results show that INV(S) has good protective effects for different types of image datasets when different advanced pretrained diffusion models are used. This study provides theoretical support and technical optimization directions for diffusion model antiattack and defense mechanisms.