腾讯朱雀实验室推出“隐形”水印，助力打击AI模型盗取行为

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"随着人工智能行业的发展，AI技术被大量应用到人们的生活中，而AI模型作为这些技术的载体，被广泛部署在云端。作为一种数字资产，AI模型面临着被窃取的风险，其安全性愈发引起业界关注。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"北京时间11月11日至12日，全球知名信息安全峰会POC 2021正式举办，腾讯朱雀实验室高级研究员Mengyun Tang和研究员Tony受邀参加，并进行了题为《Towards AI Model Security Protection（AI模型的安全保护）》的分享。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在此次分享中，腾讯朱雀实验室展示了AI模型攻防实例，并提出了一种新的模型水印生成方法，这项技术可以防御多种模型窃取方式，并且对原模型的输出几乎不产生影响，为AI模型版权提供有效的保护。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"AI模型维权，取证是难点"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"    "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"AI模型作为技术的核心载体，一旦被窃取，将可能使拥有该技术的企业或组织暴露在风险中。例如，某公司的AI模型被黑客恶意盗取后，黑客就可以复制该公司的业务，来抢占市场，获取间接经济利益，或者将模型出售给第三方，甚至勒索该公司，来获取直接经济利益。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"    "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在模型窃取方式中，代理模型攻击是一种典型的手段，它通过训练与原模型功能相似的代理模型来蒸馏原模型的知识——将原模型的输入作为其输入，原模型的输出作为其训练标签，并进行参数优化，不断拟合原模型的输出，最终达到窃取原模型知识的目的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/99\/e9\/99492175788fb80d2ba1cdc1191a99e9.jpg","alt":null,"title":"模型窃取流程","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"   "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"而在面对模型窃取攻击时，模型的原作者往往容易处于被动。因为攻击者并不直接接触原模型，所以原作者无法提供直接证据，证明被窃取的模型中含有自己的知识产权，而陷入维权困难的境地。因此，一旦模型窃取攻击泛滥，将为人工智能的发展带来更多挑战。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"“隐形”水印，AI模型版权保护新方法"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"针对上述问题，腾讯朱雀实验室结合最新的深度学习技术，推出了一套为AI模型提供保护的方法，即对疑似窃取模型进行“取证”，来证明该模型为“盗版”模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"    "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这套方法可以在预先防护阶段，生成肉眼不可见的水印，并将之添加到原模型的输出上，为原模型的输出“烙上”版权信息，同时，对原模型的输出几乎不产生影响。当AI模型被攻击时，其附带的水印也会被代理模型学习到，进而使得代理模型的输出中也含有该水印。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"    "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"随后，通过经训练的提取器，可以从代理模型的输出中精准地检测到水印的存在，并将预先嵌入的模型版权信息进行高质量的还原，从而为模型原作者提供有力的技术证据，来对抗侵权行为。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/bc\/41\/bc02a5a7c4ec76a2cc296882448e7541.jpg","alt":null,"title":"朱雀实验室提出的模型水印生成方法及其保护流程","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这项技术为AI模型提供了一道“胎记”，其意义在于，不仅能够有效地帮助AI模型作者维护自己的知识产权，还能打击“盗版”AI模型，一定程度地遏制模型窃取行为的发生，促进AI行业的生态持续健康发展。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"    "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"腾讯安全平台部下属的腾讯朱雀实验室，致力于实战级APT攻击和AI安全研究，其建设的AI安全威胁风险矩阵，专门针对人工智能行业中的潜在风险提供研究和预案，为AI业务提供安全保障。"}]}]}