扩散模型因其在图像合成等领域的卓越数据生成能力而受到广泛关注。然而,近期研究表明,扩散模型容易遭受版权侵权攻击,攻击者将经过策略性修改的非侵权图像注入训练集,诱导模型在特定中毒文本提示下生成侵权内容。为解决这一问题,我们首先提出一个防御框架 CopyrightShield,用于抵御上述攻击。具体而言,我们分析了扩散模型的记忆机制,发现攻击利用了模型对特定空间位置和提示的过拟合,使其在后门触发条件下复现中毒样本。基于此,我们提出一种利用空间掩码和数据归因的中毒样本检测方法,以量化中毒风险并准确识别隐藏的后门样本。为进一步缓解对中毒特征的记忆,我们引入一种自适应优化策略,在训练损失中引入动态惩罚项,在保持生成性能的同时减少对侵权特征的依赖。实验结果表明,CopyrightShield 在两种攻击场景下显著提升了中毒样本检测性能,平均 F1 分数达到 0.665,将首次攻击轮次(FAE)推迟了 115.2%,并将版权侵权率(CIR)降低了 56.7%。与扩散模型中目前最优的后门防御方法相比,防御效果提升约 25%,展示了其在增强扩散模型安全性方面的优越性与实用性。
英文摘要:
Diffusion models have attracted significant attention due to its exceptional data generation capabilities in fields such as image synthesis. However, recent studies have shown that diffusion models are vulnerable to copyright infringement attacks, where attackers inject strategically modified non-infringing images into the training set, inducing the model to generate infringing content under the prompt of specific poisoned captions. To address this issue, we first propose a defense framework, CopyrightShield, to defend against the above attack. Specifically, we analyze the memorization mechanism of diffusion models and find that attacks exploit the model's overfitting to specific spatial positions and prompts, causing it to reproduce poisoned samples under backdoor triggers. Based on this, we propose a poisoned sample detection method using spatial masking and data attribution to quantify poisoning risk and accurately identify hidden backdoor samples. To further mitigate memorization of poisoned features, we introduce an adaptive optimization strategy that integrates a dynamic penalty term into the training loss, reducing reliance on infringing features while preserving generative performance. Experimental results demonstrate that CopyrightShield significantly improves poisoned sample detection performance across two attack scenarios, achieving average F1-scores of 0.665, retarding the First-Attack Epoch (FAE) of 115.2% and decreasing the Copyright Infringement Rate (CIR) by 56.7%. Compared to the SoTA backdoor defense in diffusion models, the defense effect is improved by about 25%, showcasing its superiority and practicality in enhancing the security of diffusion models.
Solving Copyright Infringement on Short Video Platforms: Novel Datasets and an Audio Restoration Deep Learning Pipeline
Department of MetaBioHealth, Sungkyunkwan University, Korea(韩国成均馆大学代谢生物健康系);Department of Applied Artificial Intelligence, Sungkyunkwan University, Korea(韩国成均馆大学应用人工智能系);Department of Computer Science and Engineering, Jaume I University, Spain(西班牙伊萨贝拉大学计算机科学与工程系)
The KL3M Data Project: Copyright-Clean Training Resources for Large Language Models
Institute for the Advancement of Legal and Ethical AI (ALEA Institute)(法律与伦理AI推进研究所(ALEA研究所));CodeX—The Stanford Center for Legal Informatics(CodeX——斯坦福法律信息学中心);Illinois Tech—Chicago Kent College of Law(伊利诺伊理工大学芝加哥肯特法学院);Center for Legal Technology & Data Science, Bucerius Law School(布塞留斯法学院法律技术与数据科学中心)
CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion
Shanghai Jiao Tong University(上海交通大学);Queen’s University Belfast(贝尔法斯特女王大学);University of Southern California(南加州大学);Louisiana State University(路易斯安那州立大学)
SWA-LDM: Toward Stealthy Watermarks for Latent Diffusion Models
Software Engineering Institute, East China Normal University(华东师范大学软件工程学院);School of Computer Science and Technology, Harbin Institute of Technology (Shen Zhen)(哈尔滨工业大学(深圳)计算机科学与技术学院);College of Integrated Circuits, Zhejiang University(浙江大学集成电路学院)