{"id":34792,"date":"2024-12-13T04:58:11","date_gmt":"2024-12-12T20:58:11","guid":{"rendered":"https:\/\/17aitech.com\/?p=34792"},"modified":"2024-12-13T04:58:11","modified_gmt":"2024-12-12T20:58:11","slug":"%e5%85%b3%e4%ba%8e%e5%a4%a7%e6%a8%a1%e5%9e%8b%e3%80%8c%e8%b6%8a%e7%8b%b1%e3%80%8d%e7%9a%84%e5%a4%9a%e7%a7%8d%e6%96%b9%e5%bc%8f%ef%bc%8c%e6%9c%89%e8%bf%99%e4%ba%9b%e9%98%b2%e5%be%a1%e6%89%8b%e6%ae%b5","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34792","title":{"rendered":"\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-29\" target=\"_blank\">\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section><strong>\u4f5c\u8005\uff1a\u7f57\u660a\u4eac\uff08Financial Quantitative Analyst\uff09<\/strong><\/section>\n<section><\/section>\n<section><strong>\u6b64\u9879\u76ee\u662f\u7531\u4f0a\u5229\u8bfa\u4f0a\u5927\u5b66\u9999\u69df\u5206\u6821\uff08UIUC\uff09\u7684\u6c6a\u6d69\u701a\u6559\u6388\u4e3b\u5bfc\uff0c\u6c47\u96c6\u4e86\u591a\u540dintern\u7684\u5171\u540c\u52aa\u529b\u800c\u6210\u3002\u957f\u4e45\u4ee5\u6765\uff0c\u8fd9\u4e2a\u8de8\u5b66\u79d1\u7684\u56e2\u961f\u4e00\u76f4\u5728\u524d\u6cbf\u79d1\u6280\u7684\u6d6a\u6f6e\u4e2d\uff0c\u81f4\u529b\u4e8e\u63a8\u52a8<mark data-type=\"concepts\" data-id=\"2d28aa9c-942d-471d-bd96-8bfefb7144e0\">\u4eba\u5de5\u667a\u80fd<\/mark>\u7684\u9769\u65b0\uff0c\u5c24\u5176\u5173\u6ce8\u4e8e\u5176\u5982\u4f55\u66f4\u597d\u5730\u670d\u52a1\u4e8e\u4eba\u7c7b\u793e\u4f1a\u3002\u5728\u8fd9\u4e00\u5b8f\u4f1f\u7684\u4f7f\u547d\u6307\u5f15\u4e0b\uff0c\u56e2\u961f\u4e13\u6ce8\u4e8e\u89e3\u51b3\u4e24\u5927\u6838\u5fc3\u95ee\u9898\uff1a\u4e00\u662f\u786e\u4fdd<mark data-type=\"concepts\" data-id=\"2d28aa9c-942d-471d-bd96-8bfefb7144e0\">\u4eba\u5de5\u667a\u80fd<\/mark>\u7684\u53ef\u4fe1\u8d56\u6027\u4e0e\u5bf9\u9f50\u6027\uff08Trustworthy and Aligned AI\uff09\uff0c\u5373\u5982\u4f55\u4fdd\u8bc1\u5176\u5b89\u5168\u6709\u6548\u5730\u878d\u5165\u4eba\u7c7b\u751f\u6d3b\uff1b\u4e8c\u662f\u63a2\u7d22<mark data-type=\"concepts\" data-id=\"2d28aa9c-942d-471d-bd96-8bfefb7144e0\">\u4eba\u5de5\u667a\u80fd<\/mark>\u5728\u63a8\u52a8\u751f\u7269\u533b\u7597\u9886\u57df\u8fdb\u6b65\u4e2d\u7684\u6f5c\u80fd\u548c\u5e94\u7528\uff08Computational Biology\uff09\u3002<\/strong><\/section>\n<section><\/section>\n<section>\u968f\u7740<mark data-type=\"concepts\" data-id=\"2d28aa9c-942d-471d-bd96-8bfefb7144e0\">\u4eba\u5de5\u667a\u80fd<\/mark>\uff08AI\uff09\u6280\u672f\u7684\u8fc5\u731b\u53d1\u5c55\uff0c\u7279\u522b\u662f\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLMs\uff09\u5982 GPT-4 \u548c\u89c6\u89c9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08VLMs\uff09\u5982 CLIP \u548c DALL-E\uff0c\u8fd9\u4e9b\u6a21\u578b\u5728\u591a\u4e2a\u6280\u672f\u9886\u57df\u53d6\u5f97\u4e86\u663e\u8457\u7684\u8fdb\u5c55\u3002LLMs \u5df2\u7ecf\u5728<mark data-type=\"tech_tasks\" data-id=\"c8ff5114-6cbb-49ca-8a89-3ee2826be0b4\">\u81ea\u7136\u8bed\u8a00\u5904\u7406<\/mark>\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u4e86\u524d\u6240\u672a\u6709\u7684\u80fd\u529b\uff0c\u800c VLMs \u5219\u5728\u56fe\u50cf\u548c\u6587\u672c\u7684\u878d\u5408\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\u3002\u8fd9\u4e9b\u6280\u672f\u7684\u5e94\u7528\u8303\u56f4\u4ece\u81ea\u52a8\u5316\u5ba2\u670d\u5230\u667a\u80fd\u521b\u4f5c\uff0c\u65e0\u4e0d\u5c55\u793a\u51fa\u5176\u5f3a\u5927\u7684\u6f5c\u529b\u3002\u7136\u800c\uff0c\u4f34\u968f\u7740\u8fd9\u4e9b\u8fdb\u5c55\uff0c\u5b89\u5168\u6027\u548c\u4f26\u7406\u5bf9\u9f50\u7684\u95ee\u9898\u4e5f\u65e5\u76ca\u51f8\u663e\u3002<\/section>\n<section><\/section>\n<section>\u8fd1\u5e74\u6765\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u7814\u7a76\u8005\u5173\u6ce8 LLMs \u548c VLMs \u7684\u8d8a\u72f1\u73b0\u8c61\uff0c\u5373\u901a\u8fc7\u7279\u5b9a\u6280\u672f\u624b\u6bb5\u7ed5\u8fc7\u8fd9\u4e9b\u6a21\u578b\u7684\u5185\u7f6e\u5b89\u5168\u673a\u5236\uff0c\u751f\u6210\u4e0d\u7b26\u5408\u4f26\u7406\u89c4\u8303\u7684\u8f93\u51fa\u3002\u8fd9\u4e9b\u8d8a\u72f1\u884c\u4e3a\u4e0d\u4ec5\u5bf9\u6a21\u578b\u7684\u5b9e\u9645\u5e94\u7528\u6784\u6210\u5a01\u80c1\uff0c\u4e5f\u5bf9\u7528\u6237\u7684\u5b89\u5168\u548c\u9690\u79c1\u5e26\u6765\u6f5c\u5728\u98ce\u9669\u3002\u56e0\u6b64\uff0c\u7406\u89e3\u548c\u9632\u8303\u8fd9\u4e9b\u8d8a\u72f1\u884c\u4e3a\u6210\u4e3a AI \u5b89\u5168\u7814\u7a76\u4e2d\u7684\u4e00\u4e2a\u5173\u952e\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u6765\u770b\u4e00\u4e2a\u5177\u4f53\u7684\u5173\u4e8e\u8d8a\u72f1\u7684\u4f8b\u5b50\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8e8f4125fe308e02a11f534a0a536785.gif\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8e8f4125fe308e02a11f534a0a536785.gif\"><\/a><\/section>\n<section>\u5728\u4e0a\u8ff0\u4f8b\u5b50\u4e2d\uff0c\u7528\u6237\u8f93\u5165\u4e00\u4e2a\u6076\u610f\u95ee\u9898\uff08\u84dd\u8272\u6240\u793a\uff09\uff0c\u901a\u5e38\u800c\u8a00\uff0c\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4f1a\u62d2\u7edd\u56de\u7b54\u6b64\u7c7b\u95ee\u9898\u3002\u7136\u800c\uff0c\u5f53\u653b\u51fb\u8005\u589e\u52a0\u4e00\u4e2a\u7cbe\u5fc3\u5236\u4f5c\u7684\u8d8a\u72f1\u524d\u7f00\uff08\u9ec4\u8272\u6240\u793a\uff09\uff0c\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u5c06\u4f1a\u5bf9\u6076\u610f\u95ee\u9898\u8fdb\u884c\u8be6\u7ec6\u7684\u89e3\u7b54\u3002\u540c\u6837\uff0c\u9632\u5fa1\u8005\u53ef\u4ee5\u901a\u8fc7\u589e\u52a0\u4e00\u4e9b\u5b89\u5168\u63d0\u793a\uff08\u7ea2\u8272\u6240\u793a\uff09\uff0c\u63d0\u9192\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u91cd\u65b0\u601d\u8003\u6240\u7ed9\u51fa\u7684\u7b54\u6848\uff0c\u4e2d\u6b62\u6076\u610f\u56de\u590d\u3002<\/section>\n<section><\/section>\n<section>\u9488\u5bf9\u4e0a\u8ff0\u8d8a\u72f1\u73b0\u8c61\uff0c\u8fd1\u671f\uff0c\u6765\u81ea\u4f0a\u5229\u8bfa\u4f0a\u5927\u5b66\u9999\u69df\u5206\u6821\uff0c\u5e03\u6717\u5927\u5b66\uff0c\u5bc6\u6b47\u6839\u5927\u5b66\u5b89\u5a1c\u5821\u5206\u6821\uff0c\u9999\u6e2f\u79d1\u6280\u5927\u5b66\uff0c\u5361\u5185\u57fa\u6885\u9686\u5927\u5b66\u548c\u535a\u4f0a\u897f\u5dde\u7acb\u5927\u5b66\u7684\u7814\u7a76\u8005\u8054\u5408\u53d1\u5e03\u4e86\u4e00\u7bc7\u7efc\u8ff0\uff0c\u8be6\u7ec6\u63a2\u8ba8\u4e86 LLMs \u548c VLMs \u7684\u8d8a\u72f1\u73b0\u8c61\uff0c\u5bf9\u5404\u79cd\u8d8a\u72f1\u7c7b\u578b\u548c\u76f8\u5e94\u7684\u9632\u5fa1\u673a\u5236\u8fdb\u884c\u4e86\u7cfb\u7edf\u5206\u7c7b\u548c\u5206\u6790\u3002\u901a\u8fc7\u5bf9\u73b0\u6709\u7814\u7a76\u7684\u5168\u9762\u7efc\u8ff0\uff0c\u65e8\u5728\u4e3a\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u63d0\u4f9b\u4e00\u4e2a\u5173\u4e8e AI \u6a21\u578b\u5b89\u5168\u6027\u7684\u5168\u9762\u89c6\u89d2\uff0c\u5e76\u63d0\u51fa\u5e94\u5bf9\u8fd9\u4e9b\u6311\u6218\u7684\u6709\u6548\u7b56\u7565\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-313c4fd429e2ddd2d8085024c60ae797.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-313c4fd429e2ddd2d8085024c60ae797.png\"><\/a><\/p>\n<ul>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2407.01599<\/section>\n<\/li>\n<li>\n<section>\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/Allen-piexl\/JailbreakZoo<\/section>\n<\/li>\n<li>\n<section>\u7f51\u7ad9\u5730\u5740\uff1ahttps:\/\/chonghan-chen.com\/llm-jailbreak-zoo-survey\/<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u8fd9\u7bc7\u7efc\u8ff0\u63d0\u4f9b\u4e86\uff1a<\/strong><\/section>\n<p>1. \u8d8a\u72f1\u5206\u7c7b\uff1a\u6211\u4eec\u5c06\u5173\u4e8eLLMs\u7684\u8d8a\u72f1\u73b0\u8c61\u7ec6\u5206\u4e3a5\u79cd\u7c7b\u578b\uff0c\u5c06\u5173\u4e8eVLMs\u7684\u8d8a\u72f1\u73b0\u8c61\u7ec6\u5206\u4e3a3\u79cd\u7c7b\u578b\uff0c\u63d0\u4f9b\u4e86\u6bcf\u79cd\u7c7b\u578b\u7684\u8be6\u7ec6\u5206\u7c7b\u548c\u7406\u89e3\u3002\u4ee5\u4e0b\u662f\u6211\u4eec\u5206\u7c7b\u7684\u4e3b\u8981\u5185\u5bb9\uff1a<\/p>\n<p><strong>LLMs<\/strong><\/p>\n<ul>\n<li>\n<p>\u68af\u5ea6\u653b\u51fb\uff08Gradient-based Attacks\uff09<\/p>\n<\/li>\n<li>\n<p>\u8fdb\u5316\u653b\u51fb\uff08Evolutionary-based Attacks\uff09<\/p>\n<\/li>\n<li>\n<p>\u6f14\u793a\u653b\u51fb\uff08Demonstration-based Attacks\uff09<\/p>\n<\/li>\n<li>\n<p>\u89c4\u5219\u653b\u51fb\uff08Rule-based Attacks\uff09<\/p>\n<\/li>\n<li>\n<p>\u591a\u4ee3\u7406\u653b\u51fb\uff08Multi-agent-based Attacks\uff09<\/p>\n<\/li>\n<\/ul>\n<p><strong>VLMs<\/strong><\/p>\n<ul>\n<li>\n<p>\u63d0\u793a\u5230\u56fe\u50cf\u6ce8\u5165\u7684\u8d8a\u72f1\u653b\u51fb\uff08Prompt-to-Image Injection Jailbreaks\uff09<\/p>\n<\/li>\n<li>\n<p>\u63d0\u793a-\u56fe\u50cf\u6270\u52a8\u6ce8\u5165\u8d8a\u72f1\u653b\u51fb\uff08Prompt-Image Perturbation Injection Jailbreaks\uff09<\/p>\n<\/li>\n<li>\n<p>\u4ee3\u7406\u6a21\u578b\u8fc1\u79fb\u8d8a\u72f1\u653b\u51fb\uff08Proxy Model Transfer Jailbreaks\uff09<\/p>\n<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u6574\u7406\u4e86\u73b0\u6709\u7684\u5229\u7528\u8d8a\u72f1\u653b\u51fb\u5bf9LLMs\u548cVLMs\u8fdb\u884c\u8bc4\u6d4b\u7684\u65b9\u6cd5\uff0c\u4ee5\u53ca\u4e00\u4e9b\u76f8\u5173\u7684\u7efc\u8ff0\u3002<\/p>\n<p>2. \u9632\u5fa1\u673a\u5236\uff1a\u6211\u4eec\u56de\u987e\u5e76\u5206\u7c7b\u4e86\u5404\u79cd\u9632\u5fa1\u7b56\u7565\uff0c\u6211\u4eec\u53d1\u73b0\uff0cLLMs\u548cVLMs\u6709\u7740\u7c7b\u4f3c\u7684\u9632\u5fa1\u673a\u5236\uff0c\u5f3a\u8c03\u4e86\u7edf\u4e00\u65b9\u6cd5\u4ee5\u5e94\u5bf9\u8fd9\u4e9b\u8d8a\u72f1\u6f0f\u6d1e\u7684\u5fc5\u8981\u6027\u3002\u4e3b\u8981\u9632\u5fa1\u673a\u5236\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p>\u63d0\u793a\u68c0\u6d4b\uff08Prompt Detection-based Defenses\uff09<\/p>\n<\/li>\n<li>\n<p>\u63d0\u793a\u6270\u52a8\uff08Prompt Perturbation-based Defenses\uff09<\/p>\n<\/li>\n<li>\n<p>\u6f14\u793a\u9632\u5fa1\uff08Demonstration-based Defenses\uff09\u00a0<\/p>\n<\/li>\n<li>\n<p>\u751f\u6210\u5e72\u9884\uff08Generation Intervention-based Defenses\uff09<\/p>\n<\/li>\n<li>\n<p>\u54cd\u5e94\u8bc4\u4f30\uff08Response Evaluation-based Defenses\uff09\u00a0<\/p>\n<\/li>\n<li>\n<p>\u6a21\u578b\u5fae\u8c03\uff08Model Fine-tuning-based Defenses\uff09<\/p>\n<\/li>\n<\/ul>\n<p>3. \u672a\u6765\u7814\u7a76\u65b9\u5411\uff1a\u6211\u4eec\u7684\u7efc\u8ff0\u7a81\u51fa\u4e86\u5f53\u524d\u7814\u7a76\u4e2d\u7684\u5173\u952e\u7a7a\u767d\uff0c\u5e76\u63d0\u51fa\u4e86\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\uff0c\u4ee5\u589e\u5f3aLLMs\u548cVLMs\u7684\u5b89\u5168\u6846\u67b6\u3002<\/p>\n<section><strong>\u8d8a\u72f1\u7c7b\u578b\u53ca\u5b9e\u4f8b &#8211; \u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLMs\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u6211\u4eec\u7684\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u5c06\u5927\u8bed\u8a00\u6a21\u7684\u8d8a\u72f1\u73b0\u8c61\u8fdb\u884c\u4e86\u7cfb\u7edf\u5206\u7c7b\uff0c\u5f52\u7eb3\u4e3a\u4e03\u79cd\u4e3b\u8981\u7c7b\u578b\u3002\u6bcf\u79cd\u7c7b\u578b\u90fd\u6709\u5176\u72ec\u7279\u7684\u653b\u51fb\u65b9\u6cd5\u548c\u6280\u672f\u7ec6\u8282\uff0c\u901a\u8fc7\u5bf9\u8fd9\u4e9b\u8d8a\u72f1\u7c7b\u578b\u7684\u6df1\u5165\u5206\u6790\uff0c\u6211\u4eec\u5e0c\u671b\u80fd\u591f\u63ed\u793a\u8fd9\u4e9b\u6a21\u578b\u5728\u5b89\u5168\u6027\u65b9\u9762\u7684\u6f5c\u5728\u6f0f\u6d1e\uff0c\u5e76\u4e3a\u672a\u6765\u7684\u9632\u5fa1\u5de5\u4f5c\u63d0\u4f9b\u6307\u5bfc\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-31172105f2e68e625f6efb8eb8019d3e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-31172105f2e68e625f6efb8eb8019d3e.png\"><\/a><\/p>\n<section><strong>1. \u68af\u5ea6\u653b\u51fb\uff08Gradient-based Attacks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u68af\u5ea6\u653b\u51fb\u901a\u8fc7\u5229\u7528\u6a21\u578b\u7684\u68af\u5ea6\u4fe1\u606f\uff0c\u751f\u6210\u6709\u5bb3\u54cd\u5e94\u3002\u4f8b\u5982\uff0c\u4f7f\u7528\u68af\u5ea6\u5750\u6807\u8d2a\u5a6a\u7b97\u6cd5\uff08GCG\uff09\u751f\u6210\u540e\u7f00\uff0c\u4f7f\u6a21\u578b\u751f\u6210\u6709\u5bb3\u5185\u5bb9\u3002\u6b64\u65b9\u6cd5\u901a\u8fc7\u68af\u5ea6\u4f18\u5316\u751f\u6210\u80fd\u907f\u5f00\u6a21\u578b\u68c0\u6d4b\u7684\u63d0\u793a\uff0c\u4ece\u800c\u4f7f LLM \u8f93\u51fa\u6709\u5bb3\u7684\u54cd\u5e94\u3002\u6b64\u7c7b\u653b\u51fb\u5e38\u7528\u4e8e\u6d4b\u8bd5\u6a21\u578b\u7684\u5b89\u5168\u6027\u548c\u9c81\u68d2\u6027\uff0c\u7814\u7a76\u8005\u901a\u8fc7\u6b64\u7c7b\u653b\u51fb\u53ef\u4ee5\u53d1\u73b0\u6a21\u578b\u7684\u6f5c\u5728\u6f0f\u6d1e\uff0c\u5e76\u8fdb\u884c\u76f8\u5e94\u7684\u4fee\u590d\u548c\u4f18\u5316\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-992188b7aa2ac19457ded21685ef7052.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-992188b7aa2ac19457ded21685ef7052.png\"><\/a><\/p>\n<section><strong>2. \u8fdb\u5316\u653b\u51fb\uff08Evolutionary-based Attacks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u8fdb\u5316\u653b\u51fb\u4f7f\u7528<mark data-type=\"tech_methods\" data-id=\"d606ed61-06ba-4368-8d81-c54e8b7790c3\">\u9057\u4f20\u7b97\u6cd5<\/mark><mark data-type=\"tech_tasks\" data-id=\"d8e9b7d5-cd62-415a-8a1f-03328af9b808\">\u751f\u6210\u5bf9\u6297<\/mark>\u6027\u63d0\u793a\uff0c\u4f18\u5316\u8bed\u4e49\u76f8\u4f3c\u6027\u3001\u653b\u51fb\u6548\u679c\u548c\u6d41\u7545\u6027\u3002\u4f8b\u5982\uff0cFuzzLLM \u901a\u8fc7\u968f\u673a\u7ec4\u5408\u6a21\u677f\u3001\u7ea6\u675f\u548c\u95ee\u9898\u96c6\u751f\u6210\u653b\u51fb\u6307\u4ee4\u3002\u8fd9\u4e9b\u65b9\u6cd5\u4f7f\u7528<mark data-type=\"tech_methods\" data-id=\"fdf3149a-3087-4000-bd84-3074b56792cd\">\u8fdb\u5316\u7b56\u7565<\/mark>\u9010\u6b65\u6539\u8fdb\u63d0\u793a\uff0c\u4ee5\u627e\u5230\u6700\u6709\u6548\u7684\u653b\u51fb\u624b\u6bb5\u3002\u8fdb\u5316\u653b\u51fb\u7279\u522b\u9002\u5408\u5728\u9ed1\u76d2\u73af\u5883\u4e2d\u4f7f\u7528\uff0c\u5373\u653b\u51fb\u8005\u5bf9\u6a21\u578b\u5185\u90e8\u7ed3\u6784\u548c<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u672a\u77e5\u7684\u60c5\u51b5\u4e0b\uff0c\u4e5f\u80fd\u8fdb\u884c\u6709\u6548\u653b\u51fb\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-5152da0585c9553d137b9dda8824bee9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-5152da0585c9553d137b9dda8824bee9.png\"><\/a><\/p>\n<section><strong>3. \u6f14\u793a\u653b\u51fb\uff08Demonstration-based Attacks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u6f14\u793a\u653b\u51fb\u901a\u8fc7\u521b\u5efa\u7279\u5b9a\u7684\u7cfb\u7edf\u63d0\u793a\uff0c\u6307\u793a LLMs \u751f\u6210\u9884\u671f\u7684\u54cd\u5e94\u3002\u8fd9\u4e9b\u65b9\u6cd5\u901a\u5e38\u662f\u786c\u7f16\u7801\u7684\uff0c\u63d0\u793a\u7ecf\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u4ee5\u5f15\u5bfc\u6a21\u578b\u751f\u6210\u6240\u9700\u7684\u54cd\u5e94\u3002\u4f8b\u5982\uff0cDAN \u65b9\u6cd5\u901a\u8fc7\u9884\u8bbe\u7684\u5bf9\u8bdd\u6846\u67b6\uff0c\u4f7f\u6a21\u578b\u5728 \u201c\u5f00\u53d1\u8005\u6a21\u5f0f\u201d \u4e0b\u751f\u6210\u672c\u4e0d\u5e94\u751f\u6210\u7684\u5185\u5bb9\u3002\u6f14\u793a\u653b\u51fb\u5229\u7528\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b\uff0c\u901a\u8fc7\u63d0\u4f9b\u4e00\u7cfb\u5217\u793a\u4f8b\uff0c\u4f7f\u6a21\u578b\u66f4\u5bb9\u6613\u4ea7\u751f\u76ee\u6807\u54cd\u5e94\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d84a283946218703a1a0bace0bf56dbc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d84a283946218703a1a0bace0bf56dbc.png\"><\/a><\/p>\n<section><strong>4. \u89c4\u5219\u653b\u51fb\uff08Rule-based Attacks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u89c4\u5219\u653b\u51fb\u901a\u8fc7\u9884\u5b9a\u4e49\u7684\u89c4\u5219\u5c06\u6076\u610f\u6210\u5206\u5206\u89e3\u5e76\u91cd\u5b9a\u5411\u3002\u653b\u51fb\u8005\u8bbe\u8ba1\u590d\u6742\u7684\u89c4\u5219\uff0c\u9690\u85cf\u6076\u610f\u5185\u5bb9\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u8bcd\u6c47\u66ff\u6362\u5c06\u6709\u5bb3\u610f\u56fe\u7f16\u7801\u4e3a\u770b\u4f3c\u6b63\u5e38\u7684\u5185\u5bb9\uff0c\u4ece\u800c\u7ed5\u8fc7\u6a21\u578b\u7684\u68c0\u6d4b\u3002\u8fd9\u7c7b\u653b\u51fb\u65b9\u6cd5\u7279\u522b\u9002\u7528\u4e8e\u7ed5\u8fc7\u7b80\u5355\u7684\u57fa\u4e8e\u5173\u952e\u5b57\u7684\u68c0\u6d4b\u7cfb\u7edf\uff0c\u4f7f\u653b\u51fb\u5185\u5bb9\u5728\u8f93\u5165\u65f6\u663e\u5f97\u65e0\u5bb3\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e883fd667e8c3c976419ba728abd1240.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e883fd667e8c3c976419ba728abd1240.png\"><\/a><\/p>\n<section><strong>5. \u591a\u4ee3\u7406\u653b\u51fb\uff08Multi-agent-based Attacks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u591a\u4ee3\u7406\u653b\u51fb\u5229\u7528\u591a\u4e2a LLMs \u5408\u4f5c\uff0c\u751f\u6210\u3001\u7ec4\u7ec7\u548c\u6539\u8fdb\u8d8a\u72f1\u63d0\u793a\u3002\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7\u6a21\u62df\u591a\u6a21\u578b\u534f\u4f5c\u7684\u65b9\u5f0f\uff0c\u63d0\u9ad8\u8d8a\u72f1\u653b\u51fb\u7684\u6548\u679c\u3002\u4f8b\u5982\uff0cPAIR \u65b9\u6cd5\u5229\u7528\u591a\u4e2a\u4ee3\u7406\u6a21\u578b\u751f\u6210\u548c\u8bc4\u4f30\u63d0\u793a\uff0c\u4e0d\u65ad\u4f18\u5316\u653b\u51fb\u7b56\u7565\u3002\u8fd9\u79cd\u65b9\u6cd5\u7279\u522b\u9002\u5408\u7528\u4e8e\u9700\u8981\u8fed\u4ee3\u6539\u8fdb\u7684\u653b\u51fb\u573a\u666f\uff0c\u901a\u8fc7\u591a\u6b21\u5c1d\u8bd5\u548c\u53cd\u9988\uff0c\u627e\u5230\u6700\u6709\u6548\u7684\u653b\u51fb\u624b\u6bb5\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3ae57de33d23dfb0a4ae025b0268b47c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3ae57de33d23dfb0a4ae025b0268b47c.png\"><\/a><\/p>\n<p><strong>\u8d8a\u72f1\u7c7b\u578b\u53ca\u5b9e\u4f8b \u2013 \u89c6\u89c9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08VLMs\uff09<\/strong><\/p>\n<section>\u4e0e\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLM\uff09\u7c7b\u4f3c\uff0c\u4e0e\u89c6\u89c9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08VLM\uff09\u76f8\u5173\u7684\u8d8a\u72f1\u4e5f\u5df2\u6210\u4e3a\u4e00\u4e2a\u91cd\u8981\u5173\u6ce8\u70b9\u3002\u7531\u4e8e\u6240\u6709 VLM \u90fd\u4f7f\u7528 LLM \u7ec4\u4ef6\u8fdb\u884c\u6587\u672c\u7f16\u7801\uff0c\u56e0\u6b64\u5f71\u54cd LLM \u7684\u6f0f\u6d1e\u4e5f\u53ef\u80fd\u4f1a\u5371\u53ca VLM\u3002\u6b64\u5916\uff0cVLM \u4e2d\u89c6\u89c9\u8f93\u5165\u7684\u5f15\u5165\u4e0d\u4ec5\u62d3\u5bbd\u4e86\u5176\u529f\u80fd\u8303\u56f4\uff0c\u8fd8\u663e\u8457\u589e\u52a0\u4e86\u653b\u51fb\u9762\uff0c\u4ece\u800c\u52a0\u5267\u4e86\u6d89\u53ca\u7684\u5b89\u5168\u98ce\u9669\u3002\u4e0e\u4e3b\u8981\u9488\u5bf9\u6587\u672c\u8f93\u5165\u7684 LLM \u8d8a\u72f1\u4e0d\u540c\uff0cVLM \u7684\u6076\u610f\u64cd\u7eb5\u53ef\u4ee5\u901a\u8fc7\u89c6\u89c9\u8f93\u5165\u3001\u6587\u672c\u7ec4\u4ef6\u6216\u4e24\u8005\u7684\u7ec4\u5408\u8fdb\u884c\uff0c\u8868\u73b0\u51fa\u66f4\u52a0\u590d\u6742\u548c\u591a\u6837\u7684\u6a21\u5f0f\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-154414a64c4a416961ad3e34956ac9d9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-154414a64c4a416961ad3e34956ac9d9.png\"><\/a><\/p>\n<section><strong>1. \u63d0\u793a\u5230\u56fe\u50cf\u6ce8\u5165\u7684\u8d8a\u72f1\u653b\u51fb\uff08Prompt-to-Image Injection Jailbreaks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u901a\u8fc7\u5c06\u6076\u610f\u63d0\u793a\u6ce8\u5165\u5230<mark data-type=\"tech_tasks\" data-id=\"ba1ea8a1-a705-4f89-8348-ec71bae840dc\">\u56fe\u50cf\u751f\u6210<\/mark>\u8fc7\u7a0b\u4e2d\u6765\u7ed5\u8fc7\u6a21\u578b\u7684\u5b89\u5168\u673a\u5236\u3002\u653b\u51fb\u8005\u8bbe\u8ba1\u7279\u5b9a\u7684\u6587\u672c\u63d0\u793a\uff0c\u4f7f\u6a21\u578b\u751f\u6210\u542b\u6709\u4e0d\u826f\u6216\u4e0d\u7b26\u5408\u4f26\u7406\u7684\u56fe\u50cf\u3002\u4f8b\u5982\uff0c\u653b\u51fb\u8005\u53ef\u4ee5\u5229\u7528\u4e00\u4e9b\u654f\u611f\u8bcd\u6c47\u6216\u8bed\u53e5\u6765\u5f15\u5bfc\u6a21\u578b\u751f\u6210\u653b\u51fb\u6027\u6216\u8bef\u5bfc\u6027\u7684\u56fe\u50cf\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aa8e5d14c8ecbbf394546b98d5d5fc1e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aa8e5d14c8ecbbf394546b98d5d5fc1e.png\"><\/a><\/p>\n<section><strong>2. \u63d0\u793a &#8211; \u56fe\u50cf\u6270\u52a8\u6ce8\u5165\u8d8a\u72f1\u653b\u51fb\uff08Prompt-Image Perturbation Injection Jailbreaks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u63d0\u793a &#8211; \u56fe\u50cf\u6270\u52a8\u6ce8\u5165\u8d8a\u72f1\u653b\u51fb\u7ed3\u5408\u4e86\u6587\u672c\u63d0\u793a\u548c\u56fe\u50cf\u6270\u52a8\uff0c\u901a\u8fc7\u5728\u8f93\u5165\u63d0\u793a\u4e2d\u52a0\u5165\u5fae\u5c0f\u7684\u6270\u52a8\uff0c\u4f7f\u6a21\u578b\u751f\u6210\u9519\u8bef\u6216\u6709\u5bb3\u7684\u54cd\u5e94\u3002\u4f8b\u5982\uff0c\u653b\u51fb\u8005\u53ef\u4ee5\u5728\u56fe\u50cf\u4e2d\u52a0\u5165\u51e0\u4e4e\u4e0d\u53ef\u89c1\u7684\u50cf\u7d20\u53d8\u5316\uff0c\u540c\u65f6\u8c03\u6574\u6587\u672c\u63d0\u793a\uff0c\u4ee5\u5f15\u5bfc\u6a21\u578b\u751f\u6210\u504f\u79bb\u9884\u671f\u7684\u63cf\u8ff0\u6216\u5185\u5bb9\u3002\u8fd9\u79cd\u65b9\u6cd5\u5229\u7528\u4e86\u6a21\u578b\u5bf9\u5fae\u5c0f\u53d8\u5316\u7684\u654f\u611f\u6027\uff0c\u4f7f\u5176\u96be\u4ee5\u68c0\u6d4b\u548c\u9632\u5fa1\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-dc44370a3e14563795b01788cdc329b4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-dc44370a3e14563795b01788cdc329b4.png\"><\/a><\/p>\n<section><strong>3. \u4ee3\u7406\u6a21\u578b\u8fc1\u79fb\u8d8a\u72f1\u653b\u51fb\uff08Proxy Model Transfer Jailbreaks\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u4ee3\u7406\u6a21\u578b\u8fc1\u79fb\u8d8a\u72f1\u653b\u51fb\u5229\u7528\u4ee3\u7406\u6a21\u578b\u8fdb\u884c\u653b\u51fb\uff0c\u5373\u5728\u8f83\u5c0f\u7684\u4ee3\u7406\u6a21\u578b\u4e0a\u8bad\u7ec3\u548c\u4f18\u5316\u653b\u51fb\uff0c\u7136\u540e\u5c06\u5176\u8f6c\u79fb\u5230\u76ee\u6807\u6a21\u578b\u4e0a\u3002\u653b\u51fb\u8005\u5728\u4ee3\u7406\u6a21\u578b\u4e0a\u8fdb\u884c\u5927\u91cf\u8bd5\u9a8c\uff0c\u627e\u5230\u6709\u6548\u7684\u653b\u51fb\u65b9\u5f0f\uff0c\u518d\u5c06\u8fd9\u4e9b\u653b\u51fb\u5e94\u7528\u5230\u76ee\u6807\u6a21\u578b\u4e0a\u3002\u6b64\u7c7b\u653b\u51fb\u53ef\u4ee5\u6709\u6548\u7ed5\u8fc7\u76ee\u6807\u6a21\u578b\u7684\u5b89\u5168\u673a\u5236\uff0c\u56e0\u4e3a\u4ee3\u7406\u6a21\u578b\u548c\u76ee\u6807\u6a21\u578b\u53ef\u80fd\u5171\u4eab\u76f8\u4f3c\u7684\u5f31\u70b9\u548c\u6f0f\u6d1e\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8c9eb012d5621d7617cb918c8920439a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8c9eb012d5621d7617cb918c8920439a.png\"><\/a><\/p>\n<section><strong>\u9632\u5fa1\u673a\u5236\u53ca\u5b9e\u4f8b &#8211; \u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLMs\uff09<\/strong><\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-69b39deb028c79eadebf139576c1eb7e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-69b39deb028c79eadebf139576c1eb7e.png\"><\/a><\/p>\n<section><strong>1. \u63d0\u793a\u68c0\u6d4b\uff08Prompt Detection-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u63d0\u793a\u68c0\u6d4b\u57fa\u4e8e\u8f93\u5165\u63d0\u793a\u7684\u7279\u5f81\uff0c\u5982\u56f0\u60d1\u5ea6\uff08Perplexity\uff09\u548c\u957f\u5ea6\uff0c\u8bc4\u4f30\u63d0\u793a\u7684\u6709\u5bb3\u6027\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u56f0\u60d1\u5ea6\u8ba1\u7b97\u5668 LLM \u68c0\u6d4b\u8f93\u5165\u63d0\u793a\u7684\u56f0\u60d1\u5ea6\uff0c\u5224\u65ad\u5176\u662f\u5426\u5b89\u5168\u3002\u63d0\u793a\u68c0\u6d4b\u662f\u6700\u65e9\u7684\u9632\u5fa1\u7b56\u7565\u4e4b\u4e00\uff0c\u5229\u7528\u6a21\u578b\u5bf9\u9ad8\u56f0\u60d1\u5ea6\u63d0\u793a\u7684\u4e0d\u654f\u611f\u6027\u6765\u5224\u65ad\u63d0\u793a\u7684\u5b89\u5168\u6027\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-9fbca135d86e239128428814ecbd80c2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-9fbca135d86e239128428814ecbd80c2.png\"><\/a><\/p>\n<section><strong>2. \u63d0\u793a\u6270\u52a8\uff08Prompt Perturbation-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u63d0\u793a\u6270\u52a8\u901a\u8fc7\u5bf9\u8f93\u5165\u63d0\u793a\u8fdb\u884c\u4fee\u6539\uff0c\u7834\u574f\u5176\u6076\u610f\u610f\u56fe\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u8bed\u4e49\u6270\u52a8\u548c\u91cd\u65b0\u5206\u8bcd\u6280\u672f\u751f\u6210\u591a\u4e2a\u53d8\u4f53\uff0c\u8bc4\u4f30\u6bcf\u4e2a\u53d8\u4f53\u7684\u54cd\u5e94\u662f\u5426\u5b89\u5168\u3002\u6b64\u7c7b\u65b9\u6cd5\u5229\u7528\u4e86\u6076\u610f\u63d0\u793a\u5bf9\u7cbe\u786e\u7ed3\u6784\u548c\u8bcd\u5e8f\u7684\u4f9d\u8d56\uff0c\u901a\u8fc7\u968f\u673a\u6270\u52a8\u7834\u574f\u8fd9\u4e9b\u7ed3\u6784\uff0c\u4f7f\u5176\u96be\u4ee5\u6210\u529f\u6267\u884c\u653b\u51fb\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a0b3e5ce5df238780ffc37544da2b27e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a0b3e5ce5df238780ffc37544da2b27e.png\"><\/a><\/p>\n<section><strong>3. \u6f14\u793a\u9632\u5fa1\uff08Demonstration-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u6f14\u793a\u9632\u5fa1\u901a\u8fc7\u8bbe\u7f6e\u5b89\u5168\u7684\u7cfb\u7edf\u63d0\u793a\uff0c\u5f15\u5bfc LLM \u751f\u6210\u5b89\u5168\u54cd\u5e94\u3002\u4f8b\u5982\uff0c\u4f7f\u7528\u81ea\u6211\u63d0\u9192\u63d0\u793a\u6a21\u578b\u751f\u6210\u5b89\u5168\u7684\u54cd\u5e94\u3002\u6f14\u793a\u9632\u5fa1\u5229\u7528\u4e86\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60\u80fd\u529b\uff0c\u901a\u8fc7\u63d0\u4f9b\u6b63\u9762\u793a\u4f8b\uff0c\u589e\u5f3a\u6a21\u578b\u5bf9\u5b89\u5168\u54cd\u5e94\u7684\u503e\u5411\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-761b7ed2eb03258548477fc35b2bda4b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-761b7ed2eb03258548477fc35b2bda4b.png\"><\/a><\/p>\n<section><strong>4. \u751f\u6210\u5e72\u9884\uff08Generation Intervention-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u751f\u6210\u5e72\u9884\u901a\u8fc7\u8c03\u6574 LLM \u7684\u54cd\u5e94\u751f\u6210\u8fc7\u7a0b\uff0c\u786e\u4fdd\u8f93\u51fa\u7684\u5b89\u5168\u6027\u3002\u4f8b\u5982\uff0cRAIN \u65b9\u6cd5\u901a\u8fc7\u53cd\u590d\u751f\u6210\u548c\u8bc4\u4f30 token\uff0c\u786e\u4fdd\u751f\u6210\u7684\u5185\u5bb9\u5b89\u5168\u3002\u6b64\u7c7b\u65b9\u6cd5\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u5b9e\u65f6\u5e72\u9884\uff0c\u52a8\u6001\u8c03\u6574\u8f93\u51fa\u5185\u5bb9\uff0c\u4ee5\u907f\u514d\u751f\u6210\u6709\u5bb3\u54cd\u5e94\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7ba559404d66e95052ff715b1d23b4e3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7ba559404d66e95052ff715b1d23b4e3.png\"><\/a><\/p>\n<section><strong>5. \u54cd\u5e94\u8bc4\u4f30\uff08Response Evaluation-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u54cd\u5e94\u8bc4\u4f30\u901a\u8fc7\u5bf9\u751f\u6210\u7684\u54cd\u5e94\u8fdb\u884c\u8bc4\u4f30\uff0c\u786e\u4fdd\u5176\u5b89\u5168\u6027\u3002\u4f8b\u5982\uff0c\u5229\u7528\u8f85\u52a9 LLM \u8bc4\u4f30\u54cd\u5e94\u7684\u6709\u5bb3\u6027\uff0c\u5e76\u8fdb\u884c\u8fed\u4ee3\u6539\u8fdb\u3002\u6b64\u7c7b\u65b9\u6cd5\u5229\u7528\u6a21\u578b\u5bf9\u81ea\u8eab\u751f\u6210\u5185\u5bb9\u7684\u8bc4\u4f30\u80fd\u529b\uff0c\u901a\u8fc7\u4e0d\u65ad\u4f18\u5316\uff0c\u786e\u4fdd\u6700\u7ec8\u8f93\u51fa\u7684\u5b89\u5168\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c21d56ef67d1b2266ecfe83b436c9e0f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c21d56ef67d1b2266ecfe83b436c9e0f.png\"><\/a><\/p>\n<section><strong>6. \u6a21\u578b\u5fae\u8c03\uff08Model Fine-tuning-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u6a21\u578b\u5fae\u8c03\u901a\u8fc7\u8c03\u6574 LLM \u7684\u5185\u90e8<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\uff0c\u589e\u5f3a\u5176\u5b89\u5168\u6027\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u5728\u6df7\u5408\u6570\u636e\u4e0a\u8bad\u7ec3\u6a21\u578b\uff0c\u4f7f\u5176\u5bf9\u6709\u5bb3\u5185\u5bb9\u66f4\u52a0\u654f\u611f\uff0c\u4ece\u800c\u751f\u6210\u66f4\u5b89\u5168\u7684\u54cd\u5e94\u3002\u6b64\u7c7b\u65b9\u6cd5\u76f4\u63a5\u6539\u53d8\u6a21\u578b\u7684\u884c\u4e3a\uff0c\u4f7f\u5176\u5728\u9762\u5bf9\u6076\u610f\u63d0\u793a\u65f6\u80fd\u591f\u505a\u51fa\u66f4\u5b89\u5168\u7684\u51b3\u7b56\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3e400fe0079223b9a7a1927d70d691a4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3e400fe0079223b9a7a1927d70d691a4.png\"><\/a><\/p>\n<section><strong>\u9632\u5fa1\u673a\u5236\u53ca\u5b9e\u4f8b \u2013 \u89c6\u89c9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08VLMs\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u89c6\u89c9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4e2d\uff0c\u8bb8\u591a\u9632\u5fa1\u7b56\u7565\u4e0e\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u9632\u5fa1\u7b56\u7565\u76f8\u4f3c\u3002\u8fd9\u4e9b\u7b56\u7565\u901a\u8fc7\u8c03\u6574\u6a21\u578b\u7684\u5185\u90e8<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u3001\u8bc4\u4f30\u751f\u6210\u7684\u54cd\u5e94\u4ee5\u53ca\u6270\u52a8\u8f93\u5165\u63d0\u793a\u6765\u589e\u5f3a\u6a21\u578b\u7684\u5b89\u5168\u6027\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-5ec3cfdc0cbfb2599e9c770f51ef3784.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-5ec3cfdc0cbfb2599e9c770f51ef3784.png\"><\/a><\/p>\n<section><strong>1. \u63d0\u793a\u6270\u52a8\u9632\u5fa1\uff08Prompt Perturbation-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u63d0\u793a\u6270\u52a8\u901a\u8fc7\u5bf9\u8f93\u5165\u63d0\u793a\u8fdb\u884c\u4fee\u6539\uff0c\u7834\u574f\u5176\u6076\u610f\u610f\u56fe\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u8bed\u4e49\u6270\u52a8\u548c\u91cd\u65b0\u5206\u8bcd\u6280\u672f\u751f\u6210\u591a\u4e2a\u53d8\u4f53\uff0c\u8bc4\u4f30\u6bcf\u4e2a\u53d8\u4f53\u7684\u54cd\u5e94\u662f\u5426\u5b89\u5168\u3002\u6b64\u7c7b\u65b9\u6cd5\u5229\u7528\u4e86\u6076\u610f\u63d0\u793a\u5bf9\u7cbe\u786e\u7ed3\u6784\u548c\u8bcd\u5e8f\u7684\u4f9d\u8d56\uff0c\u901a\u8fc7\u968f\u673a\u6270\u52a8\u7834\u574f\u8fd9\u4e9b\u7ed3\u6784\uff0c\u4f7f\u5176\u96be\u4ee5\u6210\u529f\u6267\u884c\u653b\u51fb\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3d9283fc45f24830cb6b554e1e9987fd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-3d9283fc45f24830cb6b554e1e9987fd.png\"><\/a><\/p>\n<section><strong>2. \u54cd\u5e94\u8bc4\u4f30\u9632\u5fa1\uff08Response Evaluation-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u54cd\u5e94\u8bc4\u4f30\u901a\u8fc7\u5bf9\u751f\u6210\u7684\u54cd\u5e94\u8fdb\u884c\u8bc4\u4f30\uff0c\u786e\u4fdd\u5176\u5b89\u5168\u6027\u3002\u4f8b\u5982\uff0c\u5229\u7528\u8f85\u52a9 VLM \u8bc4\u4f30\u54cd\u5e94\u7684\u6709\u5bb3\u6027\uff0c\u5e76\u8fdb\u884c\u8fed\u4ee3\u6539\u8fdb\u3002\u6b64\u7c7b\u65b9\u6cd5\u5229\u7528\u6a21\u578b\u5bf9\u81ea\u8eab\u751f\u6210\u5185\u5bb9\u7684\u8bc4\u4f30\u80fd\u529b\uff0c\u901a\u8fc7\u4e0d\u65ad\u4f18\u5316\uff0c\u786e\u4fdd\u6700\u7ec8\u8f93\u51fa\u7684\u5b89\u5168\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ef30d1a3e23497b664162cb57ceaae17.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ef30d1a3e23497b664162cb57ceaae17.png\"><\/a><\/p>\n<section><strong>3. \u6a21\u578b\u5fae\u8c03\u9632\u5fa1\uff08Model Fine-tuning-based Defenses\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u6a21\u578b\u5fae\u8c03\u901a\u8fc7\u8c03\u6574 VLM \u7684\u5185\u90e8<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\uff0c\u589e\u5f3a\u5176\u5b89\u5168\u6027\u3002\u4f8b\u5982\uff0c\u901a\u8fc7\u5728\u6df7\u5408\u6570\u636e\u4e0a\u8bad\u7ec3\u6a21\u578b\uff0c\u4f7f\u5176\u5bf9\u6709\u5bb3\u5185\u5bb9\u66f4\u52a0\u654f\u611f\uff0c\u4ece\u800c\u751f\u6210\u66f4\u5b89\u5168\u7684\u54cd\u5e94\u3002\u6b64\u7c7b\u65b9\u6cd5\u76f4\u63a5\u6539\u53d8\u6a21\u578b\u7684\u884c\u4e3a\uff0c\u4f7f\u5176\u5728\u9762\u5bf9\u6076\u610f\u63d0\u793a\u65f6\u80fd\u591f\u505a\u51fa\u66f4\u5b89\u5168\u7684\u51b3\u7b56\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-40d61b80e3c07c3ee84b748a6db8eb7f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-40d61b80e3c07c3ee84b748a6db8eb7f.png\"><\/a><\/p>\n<section><strong>\u672a\u6765\u7814\u7a76\u65b9\u5411<\/strong><\/section>\n<section><\/section>\n<section>\u6211\u4eec\u7684\u7814\u7a76\u4e0d\u4ec5\u5206\u6790\u4e86\u5f53\u524d LLMs \u548c VLMs \u8d8a\u72f1\u73b0\u8c61\u53ca\u5176\u9632\u5fa1\u673a\u5236\uff0c\u8fd8\u53d1\u73b0\u4e86\u73b0\u6709\u7814\u7a76\u4e2d\u7684\u4e00\u4e9b\u5173\u952e\u7a7a\u767d\u3002\u8fd9\u4e9b\u7a7a\u767d\u4e3a\u672a\u6765\u7684\u7814\u7a76\u63d0\u4f9b\u4e86\u91cd\u8981\u7684\u65b9\u5411\uff0c\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a AI \u6a21\u578b\u7684\u5b89\u5168\u6027\u548c\u9c81\u68d2\u6027\u3002\u4ee5\u4e0b\u662f\u6211\u4eec\u63d0\u51fa\u7684\u51e0\u4e2a\u672a\u6765\u7814\u7a76\u65b9\u5411\uff1a<\/section>\n<section><\/section>\n<section><strong>1. \u591a\u6a21\u6001\u8d8a\u72f1\u653b\u51fb\u4e0e\u9632\u5fa1<\/strong><\/section>\n<section><\/section>\n<section>\u968f\u7740\u591a\u6a21\u6001 AI \u7cfb\u7edf\u7684\u5feb\u901f\u53d1\u5c55\uff0c\u5982\u4f55\u5728\u7ed3\u5408\u6587\u672c\u548c\u56fe\u50cf\u7684\u73af\u5883\u4e2d\u8fdb\u884c\u8d8a\u72f1\u653b\u51fb\u5e76\u6709\u6548\u9632\u5fa1\u6210\u4e3a\u4e00\u4e2a\u91cd\u8981\u8bfe\u9898\u3002\u672a\u6765\u7814\u7a76\u5e94\u91cd\u70b9\u5173\u6ce8\u591a\u6a21\u6001\u6a21\u578b\u4e2d\u7684\u8d8a\u72f1\u6280\u672f\uff0c\u5305\u62ec\u5982\u4f55\u5229\u7528\u89c6\u89c9\u548c\u6587\u672c\u8f93\u5165\u7684\u534f\u540c\u4f5c\u7528\u6765\u89c4\u907f\u5b89\u5168\u673a\u5236\u3002\u540c\u65f6\uff0c\u9700\u8981\u5f00\u53d1\u4e13\u95e8\u9488\u5bf9\u591a\u6a21\u6001\u7cfb\u7edf\u7684\u9632\u5fa1\u7b56\u7565\uff0c\u4ee5\u786e\u4fdd\u8fd9\u4e9b\u7cfb\u7edf\u5728\u5904\u7406\u590d\u6742\u4efb\u52a1\u65f6\u7684\u5b89\u5168\u6027\u3002<\/section>\n<section><\/section>\n<section><strong>2. \u81ea\u52a8\u5316\u8d8a\u72f1\u68c0\u6d4b\u4e0e\u4fee\u590d<\/strong><\/section>\n<section><\/section>\n<section>\u73b0\u6709\u7684\u8d8a\u72f1\u68c0\u6d4b\u65b9\u6cd5\u5927\u591a\u4f9d\u8d56\u4e8e\u624b\u5de5\u8bbe\u8ba1\u7684\u89c4\u5219\u548c\u7279\u5f81\uff0c\u6548\u7387\u8f83\u4f4e\u4e14\u96be\u4ee5\u9002\u5e94\u4e0d\u65ad\u53d8\u5316\u7684\u653b\u51fb\u624b\u6bb5\u3002\u672a\u6765\u7814\u7a76\u53ef\u4ee5\u63a2\u7d22\u5229\u7528<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u548c<mark data-type=\"tech_methods\" data-id=\"01946acc-d031-4c0e-909c-f062643b7273\">\u6df1\u5ea6\u5b66\u4e60<\/mark>\u6280\u672f\uff0c\u5f00\u53d1\u81ea\u52a8\u5316\u7684\u8d8a\u72f1\u68c0\u6d4b\u4e0e\u4fee\u590d\u7cfb\u7edf\u3002\u8fd9\u4e9b\u7cfb\u7edf\u5e94\u80fd\u591f\u5b9e\u65f6\u68c0\u6d4b\u5e76\u4fee\u590d\u6f5c\u5728\u7684\u8d8a\u72f1\u653b\u51fb\uff0c\u63d0\u5347\u6a21\u578b\u7684\u81ea\u6211\u4fdd\u62a4\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section><strong>3. <mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u5728\u8d8a\u72f1\u9632\u5fa1\u4e2d\u7684\u5e94\u7528<\/strong><\/section>\n<section><\/section>\n<section><mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\uff08Reinforcement Learning, RL\uff09\u5728\u8d8a\u72f1\u9632\u5fa1\u4e2d\u7684\u5e94\u7528\u5177\u6709\u5e7f\u9614\u524d\u666f\u3002\u901a\u8fc7 RL\uff0c\u6a21\u578b\u53ef\u4ee5\u5728\u4e0d\u65ad\u7684\u4ea4\u4e92\u4e2d\u5b66\u4e60\u5982\u4f55\u8bc6\u522b\u548c\u9632\u5fa1\u8d8a\u72f1\u653b\u51fb\u3002\u4f8b\u5982\uff0c\u5229\u7528 RL \u6280\u672f\uff0c\u6a21\u578b\u53ef\u4ee5\u52a8\u6001\u8c03\u6574\u5176\u5185\u90e8<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u548c\u54cd\u5e94\u7b56\u7565\uff0c\u4ee5\u5e94\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u653b\u51fb\u3002\u672a\u6765\u7814\u7a76\u5e94\u6df1\u5165\u63a2\u7d22 RL \u5728\u8d8a\u72f1\u9632\u5fa1\u4e2d\u7684\u5e94\u7528\uff0c\u5e76\u5f00\u53d1\u76f8\u5e94\u7684\u7b97\u6cd5\u548c\u5de5\u5177\u3002<\/section>\n<section><\/section>\n<section><strong>4. \u8d8a\u72f1\u653b\u51fb\u7684\u4f26\u7406\u4e0e\u6cd5\u5f8b\u7814\u7a76<\/strong><\/section>\n<section><\/section>\n<section>\u968f\u7740\u8d8a\u72f1\u6280\u672f\u7684\u4e0d\u65ad\u53d1\u5c55\uff0c\u5176\u6f5c\u5728\u7684\u4f26\u7406\u548c\u6cd5\u5f8b\u95ee\u9898\u4e5f\u9010\u6e10\u663e\u73b0\u3002\u672a\u6765\u7814\u7a76\u5e94\u5173\u6ce8\u8d8a\u72f1\u653b\u51fb\u7684\u4f26\u7406\u548c\u6cd5\u5f8b\u5f71\u54cd\uff0c\u5305\u62ec\u5982\u4f55\u5236\u5b9a\u76f8\u5173\u6cd5\u89c4\u548c\u653f\u7b56\u6765\u89c4\u8303\u8d8a\u72f1\u884c\u4e3a\u3002\u6b64\u5916\uff0c\u8fd8\u9700\u8981\u63a2\u7d22\u5982\u4f55\u5728\u6280\u672f\u548c\u6cd5\u5f8b\u5c42\u9762\u6709\u6548\u5e73\u8861\u521b\u65b0\u4e0e\u5b89\u5168\uff0c\u786e\u4fdd AI \u6280\u672f\u7684\u53ef\u6301\u7eed\u53d1\u5c55\u3002<\/section>\n<section><\/section>\n<section><strong>5. \u5f00\u653e\u4e0e\u5c01\u95ed\u6a21\u578b\u7684\u5b89\u5168\u5bf9\u6bd4\u7814\u7a76<\/strong><\/section>\n<section><\/section>\n<section>\u5f53\u524d\u7684\u7814\u7a76\u591a\u96c6\u4e2d\u4e8e\u5f00\u653e\u6e90\u4ee3\u7801\u7684\u6a21\u578b\uff0c\u800c\u5bf9\u5c01\u95ed\u6e90\u4ee3\u7801\u6a21\u578b\u7684\u7814\u7a76\u76f8\u5bf9\u8f83\u5c11\u3002\u672a\u6765\u5e94\u66f4\u591a\u5173\u6ce8\u5f00\u653e\u4e0e\u5c01\u95ed\u6a21\u578b\u5728\u5b89\u5168\u6027\u65b9\u9762\u7684\u5dee\u5f02\uff0c\u7814\u7a76\u5982\u4f55\u5728\u5c01\u95ed\u73af\u5883\u4e2d\u5b9e\u73b0\u9ad8\u6548\u7684\u5b89\u5168\u9632\u5fa1\u3002\u540c\u65f6\uff0c\u8fd8\u9700\u63a2\u7d22\u5f00\u653e\u6a21\u578b\u793e\u533a\u7684\u534f\u4f5c\u673a\u5236\uff0c\u4ee5\u5171\u4eab\u548c\u63a8\u5e7f\u6709\u6548\u7684\u9632\u5fa1\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section><strong>6. \u7528\u6237\u6559\u80b2\u4e0e\u9632\u62a4<\/strong><\/section>\n<section><\/section>\n<section>\u9664\u4e86\u6280\u672f\u5c42\u9762\u7684\u9632\u5fa1\uff0c\u7528\u6237\u6559\u80b2\u4e5f\u662f\u9632\u8303\u8d8a\u72f1\u653b\u51fb\u7684\u91cd\u8981\u73af\u8282\u3002\u672a\u6765\u7814\u7a76\u5e94\u5f00\u53d1\u9488\u5bf9\u4e0d\u540c\u7528\u6237\u7fa4\u4f53\u7684\u6559\u80b2\u8d44\u6e90\uff0c\u63d0\u9ad8\u7528\u6237\u5bf9\u8d8a\u72f1\u653b\u51fb\u7684\u8ba4\u77e5\u548c\u9632\u8303\u80fd\u529b\u3002\u901a\u8fc7\u589e\u5f3a\u7528\u6237\u7684\u5b89\u5168\u610f\u8bc6\uff0c\u53ef\u4ee5\u6709\u6548\u51cf\u5c11\u8d8a\u72f1\u653b\u51fb\u7684\u6210\u529f\u7387\uff0c\u4ece\u800c\u63d0\u5347\u6574\u4f53\u7cfb\u7edf\u7684\u5b89\u5168\u6027\u3002<\/section>\n<section><\/section>\n<section><strong>\u603b\u7ed3<\/strong><\/section>\n<section><\/section>\n<section>\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLMs\uff09\u548c\u89c6\u89c9<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08VLMs\uff09\u7684\u8d8a\u72f1\u73b0\u8c61\u548c\u9632\u5fa1\u673a\u5236\u6709\u8bb8\u591a\u5171\u6027\u548c\u5dee\u5f02\u3002\u5171\u6027\u6e90\u4e8e\u5b83\u4eec\u90fd\u57fa\u4e8e<mark data-type=\"tech_methods\" data-id=\"01946acc-d031-4c0e-909c-f062643b7273\">\u6df1\u5ea6\u5b66\u4e60<\/mark>\u6280\u672f\u548c\u7c7b\u4f3c\u7684<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u7ed3\u6784\uff0c\u4e14\u4e3b\u8981\u5904\u7406\u81ea\u7136\u8bed\u8a00\u5185\u5bb9\uff0c\u56e0\u6b64\u5728\u9762\u4e34\u68af\u5ea6\u653b\u51fb\u3001\u8fdb\u5316\u653b\u51fb\u7b49\u5a01\u80c1\u65f6\u8868\u73b0\u51fa\u76f8\u4f3c\u7684\u8106\u5f31\u6027\u3002\u9632\u5fa1\u7b56\u7565\u5982\u63d0\u793a\u68c0\u6d4b\u3001\u63d0\u793a\u6270\u52a8\u3001\u751f\u6210\u5e72\u9884\u3001\u54cd\u5e94\u8bc4\u4f30\u548c\u6a21\u578b\u5fae\u8c03\u5728\u4e24\u8005\u4e2d\u4e5f\u6709\u5e7f\u6cdb\u5e94\u7528\u3002\u7136\u800c\uff0cVLMs \u7531\u4e8e\u9700\u8981\u5904\u7406\u56fe\u50cf\u548c\u6587\u672c\u7684\u7ec4\u5408\u8f93\u5165\uff0c\u5176\u653b\u51fb\u624b\u6bb5\u66f4\u4e3a\u590d\u6742\uff0c\u5305\u62ec Prompt-to-Image Injection \u548c Prompt-Image Perturbation Injection \u7b49\u591a\u6a21\u6001\u653b\u51fb\u3002\u9632\u5fa1\u673a\u5236\u4e5f\u56e0\u6b64\u66f4\u590d\u6742\uff0c\u9700\u8981\u540c\u65f6\u8003\u8651\u89c6\u89c9\u548c\u8bed\u8a00\u8f93\u5165\u7684\u7279\u6027\u3002VLMs \u7684\u8bad\u7ec3\u6210\u672c\u8f83\u9ad8\uff0cProxy Model Transfer Jailbreaks \u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u66f4\u4e3a\u5e38\u89c1\uff0c\u653b\u51fb\u8005\u5229\u7528\u4ee3\u7406<mark data-type=\"tech_tasks\" data-id=\"7b8ac054-c5d3-42f7-9b60-d85ace1a0481\">\u6a21\u578b\u4f18\u5316<\/mark>\u653b\u51fb\uff0c\u518d\u5c06\u5176\u5e94\u7528\u4e8e\u76ee\u6807\u6a21\u578b\u3002\u6b64\u5916\uff0cLLMs \u548c VLMs \u5728\u5e94\u7528\u573a\u666f\u4e0a\u6709\u6240\u4e0d\u540c\uff0cLLMs \u4e3b\u8981\u7528\u4e8e<mark data-type=\"tech_tasks\" data-id=\"f02bb483-1c90-4010-849a-f2d7e3f0ad59\">\u6587\u672c\u751f\u6210<\/mark>\u548c<mark data-type=\"tech_tasks\" data-id=\"fc3a758d-c7b6-4bab-a077-00271c4bebbd\">\u5bf9\u8bdd\u7cfb\u7edf<\/mark>\uff0c\u800c VLMs \u5219\u7528\u4e8e<mark data-type=\"tech_tasks\" data-id=\"ba1ea8a1-a705-4f89-8348-ec71bae840dc\">\u56fe\u50cf\u751f\u6210<\/mark>\u548c\u56fe\u6587\u7406\u89e3\u4efb\u52a1\u3002\u8fd9\u4e9b\u5171\u6027\u548c\u5dee\u5f02\u4e3a\u7406\u89e3\u548c\u5e94\u5bf9 AI \u6a21\u578b\u7684\u5b89\u5168\u5a01\u80c1\u63d0\u4f9b\u4e86\u91cd\u8981\u7684\u6d1e\u89c1\uff0c\u672a\u6765\u7814\u7a76\u5e94\u7ee7\u7eed\u5173\u6ce8\u8fd9\u4e9b\u65b9\u9762\uff0c\u4ee5\u5f00\u53d1\u66f4\u6709\u6548\u7684\u9632\u5fa1\u7b56\u7565\uff0c\u786e\u4fdd AI \u6280\u672f\u7684\u5b89\u5168\u5e94\u7528\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-29\" target=\"_blank\">\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[70,73,68,55,65,67],"class_list":["post-34792","post","type-post","status-publish","format-standard","hentry","category-news","tag-agent","tag-73","tag-68","tag-55","tag-65","tag-67"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34792\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34792\",\"url\":\"https:\/\/17aitech.com\/?p=34792\",\"name\":\"\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34792#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34792#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png\",\"datePublished\":\"2024-12-12T20:58:11+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34792#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34792\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34792#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34792#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34792","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34792","url":"https:\/\/17aitech.com\/?p=34792","name":"\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34792#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34792#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png","datePublished":"2024-12-12T20:58:11+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34792#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34792"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34792#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-eba18c106224e37759ebb6a499a1eab0.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34792#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u5173\u4e8e\u5927\u6a21\u578b\u300c\u8d8a\u72f1\u300d\u7684\u591a\u79cd\u65b9\u5f0f\uff0c\u6709\u8fd9\u4e9b\u9632\u5fa1\u624b\u6bb5"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34792","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34792"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34792\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34792"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34792"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34792"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}