Diffusion models have demonstrated powerful performance in generating high-quality images. A typical example is text-to-image generator like Stable Diffusion. However, their widespread use also poses potential privacy risks. A key concern is membership inference attacks, which attempt to determine whether a particular data sample was used in the model training process. Existing membership inference attacks against diffusion models either directly exploit sample loss differences or rely on image-level reconstruction differences. Both approaches commonly ignore the consistency characteristics of noise prediction during the diffusion process, resulting in either low inference accuracy or high computational costs. To address these shortcomings, we propose a membership inference method based on noise aggregation analysis, and introduce a single-step, low-intensity noise injection diffusion strategy to amplify differences between member and non-member samples. Our proposed approach substantially reduces model query requirements while delivering more efficient and accurate membership inference.

翻译：扩散模型在生成高质量图像方面展现出强大性能，典型案例如文本到图像生成器Stable Diffusion。然而，其广泛应用也带来了潜在的隐私风险。其中关键问题是成员推断攻击，旨在判断特定数据样本是否用于模型训练过程。现有的面向扩散模型的成员推断攻击方法，要么直接利用样本损失差异，要么依赖图像级重建差异，这两种方式普遍忽略了扩散过程中噪声预测的一致性特征，导致推理精度较低或计算成本高昂。针对这些不足，本文提出一种基于噪声聚合分析的成员推断方法，并引入单步低强度噪声注入扩散策略，以放大成员样本与非成员样本之间的差异。所提方法大幅降低了模型查询需求，同时实现了更高效、更精准的成员推断。