针对现有无数据模型窃取攻击技术在有限的查询预算下难以拟合原始训练集分布,进而影响对目标模型决策边界拟合效果问题,提出了一种基于Stable Diffusion的模型窃取攻击方法(Model Extraction Attack Based on Stable Diffusion,MEASD)....针对现有无数据模型窃取攻击技术在有限的查询预算下难以拟合原始训练集分布,进而影响对目标模型决策边界拟合效果问题,提出了一种基于Stable Diffusion的模型窃取攻击方法(Model Extraction Attack Based on Stable Diffusion,MEASD).利用预训练的Stable Diffusion生成训练数据可能涵盖多个域,并包含大量非判别性样本,设计了ILAF方法以优化Stable Diffusion生成的数据品质.将高质量合成数据的原始样本与由对抗样本生成器生成的对抗样本组成替代训练集.由DPA模块组合的替代模型基于替代训练集拟合目标模型的决策边界.实验结果表明,在四个主流的基准数据集上与EBFA和DMEAE方法相比,所提的MEASD方法能够将目标模型决策边界的拟合程度提高至84%,对目标模型的黑盒对抗攻击成功率超过68%,并且查询预算较低.MEASD方法能够有效地提升目标模型决策边界拟合效果及攻击成功率.展开更多
Digital rock analysis(DRA)is fundamental for geo-energy research,enabling the characterisation of microstructures for applications like hydrocarbon recovery,carbon storage,and groundwater modelling.Although 2D CT imag...Digital rock analysis(DRA)is fundamental for geo-energy research,enabling the characterisation of microstructures for applications like hydrocarbon recovery,carbon storage,and groundwater modelling.Although 2D CT images provide valuable pore-scale data,the scarcity of real-world datasets limits the effectiveness of advanced analysis.Generative AI presents a promising approach for synthesizing high-quality rock images but faces key challenges,including high computational demands,insufficient evaluation metrics,and the trade-off between image fidelity and diversity.To address these limitations,this study proposes the use of Low-Rank Adaptation(LoRA)for fine-tuning stable diffusion models,significantly reducing computational requirements while maintaining image quality.A systematic investigation was conducted to evaluate the influence of LoRA training parameters,including rank and learning rate,on the quality of generated images.Image outputs were assessed using both standard generative metrics,such as Kernel Inception Distance(KID),and domain-specific metrics,including porosity,pore count,and pore area distributions.The optimised LoRA-enhanced diffusion model achieved a 92.6% reduction in KID relative to baseline models,while also improving inference speed.Building on these advancements,this study demonstrates that the LoRA-enhanced diffusion model significantly improves neural network extrapolation in incomplete data scenarios through statistically consistent synthetic generation.Despite control challenges,this approach reduces costs and enables diverse applications,bridging fundamental rock physics with practical energy research.展开更多
文摘针对现有无数据模型窃取攻击技术在有限的查询预算下难以拟合原始训练集分布,进而影响对目标模型决策边界拟合效果问题,提出了一种基于Stable Diffusion的模型窃取攻击方法(Model Extraction Attack Based on Stable Diffusion,MEASD).利用预训练的Stable Diffusion生成训练数据可能涵盖多个域,并包含大量非判别性样本,设计了ILAF方法以优化Stable Diffusion生成的数据品质.将高质量合成数据的原始样本与由对抗样本生成器生成的对抗样本组成替代训练集.由DPA模块组合的替代模型基于替代训练集拟合目标模型的决策边界.实验结果表明,在四个主流的基准数据集上与EBFA和DMEAE方法相比,所提的MEASD方法能够将目标模型决策边界的拟合程度提高至84%,对目标模型的黑盒对抗攻击成功率超过68%,并且查询预算较低.MEASD方法能够有效地提升目标模型决策边界拟合效果及攻击成功率.
基金funding from Innovate UK(reference number:10003208)the China Scholarship Council(Grant No.CSC 202408420030).
文摘Digital rock analysis(DRA)is fundamental for geo-energy research,enabling the characterisation of microstructures for applications like hydrocarbon recovery,carbon storage,and groundwater modelling.Although 2D CT images provide valuable pore-scale data,the scarcity of real-world datasets limits the effectiveness of advanced analysis.Generative AI presents a promising approach for synthesizing high-quality rock images but faces key challenges,including high computational demands,insufficient evaluation metrics,and the trade-off between image fidelity and diversity.To address these limitations,this study proposes the use of Low-Rank Adaptation(LoRA)for fine-tuning stable diffusion models,significantly reducing computational requirements while maintaining image quality.A systematic investigation was conducted to evaluate the influence of LoRA training parameters,including rank and learning rate,on the quality of generated images.Image outputs were assessed using both standard generative metrics,such as Kernel Inception Distance(KID),and domain-specific metrics,including porosity,pore count,and pore area distributions.The optimised LoRA-enhanced diffusion model achieved a 92.6% reduction in KID relative to baseline models,while also improving inference speed.Building on these advancements,this study demonstrates that the LoRA-enhanced diffusion model significantly improves neural network extrapolation in incomplete data scenarios through statistically consistent synthetic generation.Despite control challenges,this approach reduces costs and enables diverse applications,bridging fundamental rock physics with practical energy research.
