文生视频大模型设计的安全风险及其矫治
Security Risks and Remediation of the Design of Text-to-Video Generative Models
DOI: 10.12677/design.2024.96672, PDF,   
作者: 陈 钲, 陈 靖:南京林业大学,人文社会科学学院,江苏 南京
关键词: 文生视频大模型安全风险Text-to-Video Generative Models Security Risks
摘要: 本文深入探讨了文生视频大模型设计中的安全风险及其矫治策略。随着人工智能技术的快速发展,文生视频大模型如Sora和PixelDance等,已经能够根据文本描述生成视频内容,为影视、广告、教育等行业带来了革命性的变化。然而,这些技术进步也伴随着隐私泄露、数据安全、道德价值偏离等安全风险。本文分析了训练数据、提示词注入攻击、电信欺诈、道德价值偏离和人机交互等方面的风险,并介绍了差分隐私和联邦学习等风险治理策略。
Abstract: This article delves into the safety risks and rectification strategies in the design of text-to-video generative models. With the rapid advancement of AI technology, models such as Sora and PixelDance can generate video content based on textual descriptions, revolutionizing industries like film, advertising, and education. However, these technological leaps also come with safety risks, including privacy breaches, data security, and moral value deviations. The article analyzes risks in training data, prompt injection attacks, telecommunication fraud, moral value shifts, and human-computer interaction, and proposes risk governance strategies like differential privacy and federated learning to ensure the healthy development of technology and the harmonious stability of society.
文章引用:陈钲, 陈靖. 文生视频大模型设计的安全风险及其矫治[J]. 设计进展, 2024, 9(6): 109-115. https://doi.org/10.12677/design.2024.96672

参考文献

[1] 本刊综合. 全国两会启航安全保密新征程[J]. 保密工作, 2024(3): 5-7.
[2] 腾讯研究院. 大模型安全与伦理研究报告[EB/OL].
https://www.tisi.org/27403/, 2024-09-05.
[3] 朱光辉, 王喜文. ChatGPT的运行模式、关键技术及未来图景[J]. 新疆师范大学学报(哲学社会科学版), 2023, 44(4): 113-122.
[4] Pan, Y., Mei, T., Yao, T., Li, H. and Rui, Y. (2016) Jointly Modeling Embedding and Translation to Bridge Video and Language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, 27-30 June 2016, 4594-4602. [Google Scholar] [CrossRef
[5] Tulyakov, S., Liu, M., Yang, X. and Kautz, J. (2018) MoCoGAN: Decomposing Motion and Content for Video Generation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, 18-23 June 2018, 526-1535. [Google Scholar] [CrossRef
[6] Wei, A., Haghtalab, N. and Steinhardt, J. (2024) Jailbroken: How Does LLM Safety Training Fail? arXiv: 2307.02483.
[7] Sun, H., Zhang, Z., Deng, J., et al. (2023) Safety Assessment of Chinese Large Language Models. arXiv: 2304. 10436.
[8] Liu ,Y., Deng, G., Xu, Z., et al. (2023) Jailbreaking ChatGPT via Prompt Engineering: An Empirical Study. arXiv: 2305. 13860.
[9] 胡泳, 刘纯懿. 大语言模型“数据为王”: 训练数据的价值、迷思与数字传播的未来挑战[J]. 西北师大学报(社会科学版), 2024, 61(3): 43-54.
[10] 许雪晨. ChatGPT等大语言模型赋能数字时代金融业: 基于隐私保护, 算法歧视与系统风险[J]. 暨南学报(哲学社会科学版), 2024, 46(8): 108-122.
[11] Yang, H., Ma, X., Du, K., Li, Z., Duan, H., Su, X., et al. (2017) How to Learn Klingon without a Dictionary: Detection and Measurement of Black Keywords Used by the Underground Economy. 2017 IEEE Symposium on Security and Privacy (SP), San Jose, 22-26 May 2017, 751-769. [Google Scholar] [CrossRef
[12] Yang, M., Guo, T., Zhu, T., Tjuawinata, I., Zhao, J. and Lam, K. (2024) Local Differential Privacy and Its Applications: A Comprehensive Survey. Computer Standards & Interfaces, 89, Article ID: 103827. [Google Scholar] [CrossRef
[13] Wen, J., Zhang, Z., Lan, Y., Cui, Z., Cai, J. and Zhang, W. (2022) A Survey on Federated Learning: Challenges and Applications. International Journal of Machine Learning and Cybernetics, 14, 513-535. [Google Scholar] [CrossRef] [PubMed]