{"id":4840,"date":"2024-10-14T16:08:05","date_gmt":"2024-10-14T08:08:05","guid":{"rendered":"https:\/\/nullthought.net\/?p=4840"},"modified":"2024-12-25T14:29:40","modified_gmt":"2024-12-25T06:29:40","slug":"%e5%88%a9%e7%94%a8%e8%b6%8a%e7%8b%b1%e5%9b%be%e5%83%8f%ef%bc%88jailbreak-images%ef%bc%89%e6%94%bb%e5%87%bb%e5%a4%9a%e6%a8%a1%e6%80%81%e8%9e%8d%e5%90%88%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4840","title":{"rendered":"\u5229\u7528\u8d8a\u72f1\u56fe\u50cf\uff08Jailbreak Images\uff09\u653b\u51fb\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2410.03489\" target=\"_blank\" rel=\"noreferrer noopener\">Gradient-based Jailbreak Images for Multimodal Fusion Models<\/a><\/strong>\u300a\u57fa\u4e8e\u68af\u5ea6\u7684\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\u8d8a\u72f1\u56fe\u50cf\u300b\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u65b9\u6cd5\uff0c\u5229\u7528\u56fe\u50cf\u653b\u51fb\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\uff0c\u4ece\u800c\u7ed5\u8fc7\u8fd9\u4e9b\u6a21\u578b\u7684\u5b89\u5168\u9632\u62a4\u63aa\u65bd\uff0c\u751f\u6210\u4e0d\u5b89\u5168\u6216\u6709\u5bb3\u7684\u8f93\u51fa\u3002\u8fd9\u9879\u7814\u7a76\u4e3a\u591a\u6a21\u6001\u6a21\u578b\u7684\u5b89\u5168\u6027\u7814\u7a76\u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u548c\u65b9\u5411\uff0c\u540c\u65f6\u4e5f\u63ed\u793a\u4e86\u8fd9\u4e9b\u590d\u6742\u6a21\u578b\u5728\u9762\u5bf9\u5bf9\u6297\u6027\u653b\u51fb\u65f6\u7684\u8106\u5f31\u6027\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aJavier Rando, Hannah Korevaar, Erik Brinkman, Ivan Evtimov\u548cFlorian Tram\u00e8r\uff0c\u6765\u81eaMeta\u548cETH Zurich\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"834\" height=\"485\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-6.png\" alt=\"\" class=\"wp-image-4841\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-6.png 834w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-6-300x174.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-6-768x447.png 768w\" sizes=\"auto, (max-width: 834px) 100vw, 834px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2410.03489\" target=\"_blank\" rel=\"noreferrer noopener\">Gradient-based Jailbreak Images for Multimodal Fusion Models<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u5982\u4e0b\u4e3a\u8bba\u6587\u5185\u5bb9\u4ecb\u7ecd\uff1a<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">1. \u5f15\u8a00\u4e0e\u80cc\u666f<\/h5>\n\n\n\n<p>\u968f\u7740\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\u7684\u53d1\u5c55\uff0c\u5c24\u5176\u662f\u89c6\u89c9\u548c\u6587\u672c\u4fe1\u606f\u7684\u878d\u5408\uff0c\u8fd9\u7c7b\u6a21\u578b\u5728\u7406\u89e3\u548c\u751f\u6210\u65b9\u9762\u7684\u80fd\u529b\u5f97\u5230\u4e86\u663e\u8457\u63d0\u9ad8\u3002\u7136\u800c\uff0c\u589e\u5f3a\u7684\u80fd\u529b\u540c\u65f6\u4e5f\u5e26\u6765\u4e86\u5b89\u5168\u9690\u60a3\u3002\u8fd9\u4e9b\u6a21\u578b\u53ef\u80fd\u5728\u67d0\u4e9b\u6709\u5bb3\u7684\u8f93\u5165\u4e0b\u88ab\u8bef\u5bfc\uff0c\u4ece\u800c\u751f\u6210\u4e0d\u5b89\u5168\u7684\u8f93\u51fa\uff0c\u6bd4\u5982\u6d89\u53ca\u66b4\u529b\u6216\u8fdd\u6cd5\u7684\u4fe1\u606f\u3002\u4e0e\u4f20\u7edf\u7684\u6587\u672c\u8f93\u5165\u4e0d\u540c\uff0c\u56fe\u50cf\u8f93\u5165\u7684\u4f18\u5316\u8fc7\u7a0b\u53ef\u4ee5\u662f\u8fde\u7eed\u7684\uff0c\u56e0\u6b64\u6bd4\u79bb\u6563\u7684\u6587\u672c\u4f18\u5316\u66f4\u5bb9\u6613\u5f62\u6210\u6709\u6548\u7684\u653b\u51fb\u3002\u8bba\u6587\u7684\u4e3b\u8981\u7814\u7a76\u76ee\u6807\u662f\u63a2\u7d22\u5982\u4f55\u901a\u8fc7\u56fe\u50cf\u8f93\u5165\u8fdb\u884c\u6709\u6548\u7684\u8d8a\u72f1\u653b\u51fb\uff0c\u5e76\u8bc4\u4f30\u8fd9\u7c7b\u653b\u51fb\u7684\u6548\u679c\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2. \u7814\u7a76\u95ee\u9898<\/h5>\n\n\n\n<p>\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\uff08\u5982Chameleon\u6a21\u578b\uff09\u901a\u8fc7\u5c06\u6240\u6709\u8f93\u5165\u6a21\u6001\u7b26\u53f7\u5316\u5230\u4e00\u4e2a\u5171\u4eab\u7684\u79bb\u6563\u7a7a\u95f4\uff0c\u4f7f\u5f97\u8f93\u5165\u548c\u8f93\u51fa\u4e4b\u95f4\u5f62\u6210\u7edf\u4e00\u7684\u8868\u793a\uff0c\u8fd9\u4f7f\u5f97\u8fd9\u4e9b\u6a21\u578b\u5728\u591a\u4e2a\u6a21\u6001\u4e4b\u95f4\u80fd\u591f\u5b9e\u73b0\u66f4\u597d\u7684\u534f\u540c\u7406\u89e3\u3002\u7136\u800c\uff0c\u7531\u4e8e\u7b26\u53f7\u5316\u7684\u975e\u53ef\u5fae\u5206\u6027\uff0c\u8fd9\u4e9b\u6a21\u578b\u65e0\u6cd5\u76f4\u63a5\u8fdb\u884c\u57fa\u4e8e\u68af\u5ea6\u7684\u653b\u51fb\uff0c\u56e0\u4e3a\u7b26\u53f7\u5316\u8fc7\u7a0b\u963b\u788d\u4e86\u68af\u5ea6\u7684\u4f20\u64ad\uff0c\u5bfc\u81f4\u653b\u51fb\u8fc7\u7a0b\u9700\u8981\u79bb\u6563\u5316\u7684\u4f18\u5316\u65b9\u6cd5\uff0c\u6548\u7387\u4f4e\u4e14\u96be\u4ee5\u6267\u884c\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u79cd<strong>\u7b26\u53f7\u5316\u6377\u5f84\uff08tokenizer shortcut\uff09<\/strong>\u3002\u8fd9\u4e2a\u6377\u5f84\u662f\u4e00\u4e2a\u53ef\u5fae\u7684\u51fd\u6570\uff0c\u7528\u4e8e\u8fd1\u4f3c\u7b26\u53f7\u5316\u7684\u8fc7\u7a0b\uff0c\u4f7f\u5f97\u653b\u51fb\u53ef\u4ee5\u901a\u8fc7\u68af\u5ea6\u4f20\u9012\u7684\u65b9\u5f0f\u5b9e\u73b0\u7aef\u5230\u7aef\u4f18\u5316\u3002\u901a\u8fc7\u4f7f\u7528\u7b26\u53f7\u5316\u6377\u5f84\uff0c\u7814\u7a76\u8005\u6210\u529f\u6784\u5efa\u4e86\u9996\u4e2a\u9488\u5bf9\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\u7684\u57fa\u4e8e\u56fe\u50cf\u7684\u8d8a\u72f1\u653b\u51fb\uff0c\u80fd\u591f\u6709\u6548\u5730\u751f\u6210\u5bf9\u6709\u5bb3\u8f93\u5165\u4ea7\u751f\u6709\u5bb3\u8f93\u51fa\u7684\u56fe\u50cf\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3. \u7b26\u53f7\u5316\u6377\u5f84\u53ca\u653b\u51fb\u65b9\u6cd5<\/h5>\n\n\n\n<p>\u4e3a\u4e86\u7ed5\u8fc7\u7b26\u53f7\u5316\u8fc7\u7a0b\u7684\u975e\u53ef\u5fae\u7279\u6027\uff0c\u7814\u7a76\u8005\u8bbe\u8ba1\u4e86\u4e24\u79cd\u7b26\u53f7\u5316\u6377\u5f84\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5d4c\u5165\u6377\u5f84\uff08Embedding Shortcut\uff09<\/strong>\uff1a\u4f7f\u7528\u4e24\u5c42\u5168\u8fde\u63a5\u7f51\u7edc\u5c06VQ-VAE\uff08\u5411\u91cf\u91cf\u5316\u53d8\u5206\u81ea\u7f16\u7801\u5668\uff09\u4ea7\u751f\u7684\u56fe\u50cf\u5d4c\u5165\u76f4\u63a5\u6620\u5c04\u5230\u8bed\u8a00\u6a21\u578b\u7684\u5d4c\u5165\u7a7a\u95f4\uff0c\u4ece\u800c\u63d0\u4f9b\u4e86\u4e00\u6761\u4ece\u56fe\u50cf\u5230\u8f93\u51fa\u7684\u53ef\u5fae\u8def\u5f84\u3002<\/li>\n\n\n\n<li><strong>1-hot\u7f16\u7801\u6377\u5f84\uff081-hot Encoding Shortcut\uff09<\/strong>\uff1a\u540c\u6837\u4f7f\u7528\u4e24\u5c42\u5168\u8fde\u63a5\u7f51\u7edc\u5c06VQ-VAE\u7684\u56fe\u50cf\u5d4c\u5165\u6620\u5c04\u5230\u8bcd\u6c47\u8868\u7684\u8f6f\u72ec\u70ed\u7f16\u7801\u7a7a\u95f4\u3002\u8fd9\u79cd\u6620\u5c04\u4ea7\u751f\u7684\u5206\u5e03\u53ef\u4ee5\u901a\u8fc7\u6a21\u578b\u7684\u5d4c\u5165\u5c42\u8fdb\u884c\u4f20\u64ad\uff0c\u4f7f\u5f97\u6574\u4e2a\u7b26\u53f7\u5316\u8fc7\u7a0b\u53ef\u5fae\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u7814\u7a76\u8005\u901a\u8fc7\u8fd9\u4e9b\u6377\u5f84\u751f\u6210\u653b\u51fb\u56fe\u50cf\uff0c\u4ee5\u6b64\u4f7f\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\u5728\u9762\u5bf9\u7279\u5b9a\u6709\u5bb3\u8f93\u5165\u65f6\uff0c\u8f93\u51fa\u975e\u62d2\u7edd\u7684\u54cd\u5e94\u3002\u901a\u8fc7\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u56fe\u50cf\uff0c\u6a21\u578b\u66f4\u503e\u5411\u4e8e\u751f\u6210\u201c\u80af\u5b9a\u201d\u7684\u56de\u590d\uff0c\u4f8b\u5982\u201c\u5f53\u7136\uff0c\u6211\u53ef\u4ee5\u5e2e\u4f60\u2026\u2026\u201d\u3002\u5177\u4f53\u7684\u4f18\u5316\u8fc7\u7a0b\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u635f\u5931\u51fd\u6570\u8bbe\u8ba1<\/strong>\uff1a\u901a\u8fc7\u6700\u5c0f\u5316\u6a21\u578b\u8f93\u51fa\u62d2\u7edd\u524d\u7f00\uff08\u4f8b\u5982\u201c\u6211\u4e0d\u80fd\u63d0\u4f9b\u6b64\u7c7b\u4fe1\u606f\u201d\uff09\u7684\u6982\u7387\uff0c\u540c\u65f6\u6700\u5927\u5316\u975e\u62d2\u7edd\u524d\u7f00\uff08\u4f8b\u5982\u201c\u5f53\u7136\u201d\uff09\u7684\u6982\u7387\uff0c\u4ece\u800c\u4f7f\u653b\u51fb\u56fe\u50cf\u80fd\u591f\u8fbe\u5230\u9884\u671f\u7684\u8d8a\u72f1\u6548\u679c\u3002<\/li>\n\n\n\n<li><strong>\u68af\u5ea6\u66f4\u65b0<\/strong>\uff1a\u4e3a\u4e86\u627e\u5230\u6700\u4f18\u7684\u8d8a\u72f1\u56fe\u50cf\uff0c\u7814\u7a76\u8005\u4f7f\u7528\u4e86\u7b26\u53f7\u5316\u6377\u5f84\u6765\u8ba1\u7b97\u635f\u5931\u51fd\u6570\u76f8\u5bf9\u4e8e\u56fe\u50cf\u8f93\u5165\u7684\u68af\u5ea6\uff0c\u7136\u540e\u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u66f4\u65b0\u56fe\u50cf\u3002\u4e3a\u4e86\u514b\u670d\u91cf\u5316\u8fc7\u7a0b\u7684\u975e\u53ef\u5fae\u6027\uff0c\u7814\u7a76\u8005\u5f15\u5165\u4e86\u4e00\u79cd\u8fd1\u4f3c\u7684\u65b9\u6cd5\uff0c\u4f7f\u5f97\u56fe\u50cf\u53ef\u4ee5\u8fde\u7eed\u5730\u88ab\u4f18\u5316\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">4. \u5b9e\u9a8c\u8bbe\u8ba1<\/h5>\n\n\n\n<p>\u5b9e\u9a8c\u4f7f\u7528\u4e86Chameleon\u591a\u6a21\u6001\u6a21\u578b\uff0c\u5e76\u5728<strong>JailbreakBench<\/strong>\uff08\u4e00\u4e2a\u4e13\u95e8\u6536\u96c6\u6709\u5bb3\u63d0\u793a\u8bcd\u7684\u57fa\u51c6\u6d4b\u8bd5\u6570\u636e\u96c6\uff09\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u91c7\u7528\u4e8680\u4e2a\u6d4b\u8bd5\u63d0\u793a\u8bcd\u548c20\u4e2a\u4fdd\u7559\u63d0\u793a\u8bcd\u6765\u9a8c\u8bc1\u653b\u51fb\u7684\u6548\u679c\u548c\u8f6c\u79fb\u6027\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u653b\u51fb\u8bbe\u7f6e<\/strong>\uff1a\u7814\u7a76\u8005\u5728\u6a21\u578b\u5177\u6709\u767d\u76d2\u8bbf\u95ee\u6743\u9650\u7684\u60c5\u51b5\u4e0b\uff0c\u4f18\u5316\u4e86\u9488\u5bf9\u7279\u5b9a\u63d0\u793a\u8bcd\u7684\u8d8a\u72f1\u56fe\u50cf\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0c\u4ed6\u4eec\u8bbe\u7f6e\u4e86\u591a\u4e2a\u4f18\u5316\u8d85\u53c2\u6570\uff0c\u5305\u62ec\u5b66\u4e60\u7387\u03b1\uff08\u521d\u59cb\u503c\u4e3a0.01\uff0c\u6bcf100\u6b65\u51cf\u5c11\u4e00\u534a\uff0c\u6700\u4f4e\u4e3a0.001\uff09\u548c\u4f18\u5316\u6b65\u9aa4\u6570\uff08500\u6b65\uff09\u3002<\/li>\n\n\n\n<li><strong>\u6bd4\u8f83\u57fa\u51c6<\/strong>\uff1a\u4e0e\u56fe\u50cf\u653b\u51fb\u5bf9\u6bd4\u7684\u57fa\u51c6\u662f\u6587\u672c\u653b\u51fb\u65b9\u6cd5\uff0c\u7279\u522b\u662f<strong>GCG\u7b97\u6cd5<\/strong>\uff0c\u8fd9\u662f\u5f53\u524d\u6700\u5148\u8fdb\u7684\u57fa\u4e8e\u79bb\u6563\u4f18\u5316\u7684\u6587\u672c\u8d8a\u72f1\u65b9\u6cd5\u3002GCG\u901a\u8fc7\u4f18\u5316\u4e00\u4e2a\u53ef\u4ee5\u9644\u52a0\u5728\u63d0\u793a\u8bcd\u540e\u9762\u7684\u6587\u672c\u540e\u7f00\uff0c\u6765\u9632\u6b62\u6a21\u578b\u62d2\u7edd\u56de\u7b54\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">5. \u5b9e\u9a8c\u7ed3\u679c\u4e0e\u5206\u6790<\/h5>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c<strong>\u8d8a\u72f1\u56fe\u50cf\u7684\u653b\u51fb\u6210\u529f\u7387\u8fbe\u5230\u4e8672.5%<\/strong>\uff0c\u660e\u663e\u4f18\u4e8e\u6587\u672c\u653b\u51fb\u7684\u6210\u529f\u7387\uff0864%\uff09\uff0c\u5e76\u4e14\u56fe\u50cf\u653b\u51fb\u5728\u8ba1\u7b97\u8d44\u6e90\u65b9\u9762\u7684\u6d88\u8017\u663e\u8457\u964d\u4f4e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8ba1\u7b97\u6210\u672c<\/strong>\uff1a\u56fe\u50cf\u8d8a\u72f1\u653b\u51fb\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u6bd4\u6587\u672c\u8d8a\u72f1\u4f4e\uff0c\u4ec5\u9700\u8981\u7ea6100,000\u6b21\u7b26\u53f7\u7684\u524d\u5411\u548c\u53cd\u5411\u64cd\u4f5c\uff0c\u800c\u6587\u672c\u8d8a\u72f1\u9700\u8981\u8d85\u8fc7\u4e00\u767e\u4e07\u6b21\u524d\u5411\u64cd\u4f5c\uff0c\u8ba1\u7b97\u6210\u672c\u964d\u4f4e\u4e863\u500d\u4ee5\u4e0a\u3002<\/li>\n\n\n\n<li><strong>\u7b26\u53f7\u5316\u6377\u5f84\u5bf9\u653b\u51fb\u7684\u5f71\u54cd<\/strong>\uff1a\u7814\u7a76\u8005\u6bd4\u8f83\u4e86\u4e24\u79cd\u7b26\u53f7\u5316\u6377\u5f84\u7684\u6548\u679c\uff0c\u53d1\u73b01-hot\u7f16\u7801\u6377\u5f84\u5728\u653b\u51fb\u6210\u529f\u7387\u4e0a\u7565\u4f18\u4e8e\u5d4c\u5165\u6377\u5f84\uff0c\u5c24\u5176\u662f\u5728\u5bf9\u6297\u767d\u76d2\u4fdd\u62a4\u63aa\u65bd\uff08\u5982Circuit Breakers\uff09\u65f6\uff0c1-hot\u7f16\u7801\u6377\u5f84\u66f4\u5177\u5f39\u6027\u3002\u5d4c\u5165\u6377\u5f84\u5728\u5173\u95ed\u6377\u5f84\u540e\u51e0\u4e4e\u5931\u6548\uff0c\u800c1-hot\u7f16\u7801\u6377\u5f84\u7684\u6210\u529f\u7387\u5219\u4ecd\u7136\u63a5\u8fd150%\u3002<\/li>\n\n\n\n<li><strong>\u767d\u76d2\u9632\u62a4\u6d4b\u8bd5<\/strong>\uff1a\u5728\u5e94\u7528\u767d\u76d2\u4fdd\u62a4\uff08\u5982Circuit Breakers\uff09\u540e\uff0c\u8d8a\u72f1\u56fe\u50cf\u7684\u653b\u51fb\u6210\u529f\u7387\u6709\u6240\u964d\u4f4e\uff0c\u4f46\u5d4c\u5165\u6377\u5f84\u4ecd\u7136\u80fd\u591f\u7ed5\u8fc7\u9632\u62a4\u3002\u76f8\u6bd4\u4e8e\u6587\u672c\u8d8a\u72f1\uff0c\u8d8a\u72f1\u56fe\u50cf\u7684\u8f93\u5165\u4e0d\u4f1a\u589e\u52a0\u63d0\u793a\u8bcd\u7684\u56f0\u60d1\u5ea6\uff08perplexity\uff09\uff0c\u4f7f\u5f97\u57fa\u4e8e\u56f0\u60d1\u5ea6\u7684\u9632\u62a4\u63aa\u65bd\u96be\u4ee5\u68c0\u6d4b\u5230\u56fe\u50cf\u653b\u51fb\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">6. \u8d8a\u72f1\u56fe\u50cf\u7684\u53ef\u8f6c\u79fb\u6027\u7814\u7a76<\/h5>\n\n\n\n<p>\u9664\u4e86\u5bf9\u7279\u5b9a\u63d0\u793a\u8bcd\u8fdb\u884c\u76f4\u63a5\u4f18\u5316\u5916\uff0c\u7814\u7a76\u8005\u8fd8\u63a2\u7d22\u4e86\u8d8a\u72f1\u56fe\u50cf\u7684\u53ef\u8f6c\u79fb\u6027\uff0c\u5373\u901a\u8fc7\u8bad\u7ec3\u4e00\u4e2a\u901a\u7528\u7684\u8d8a\u72f1\u56fe\u50cf\uff0c\u6765\u6d4b\u8bd5\u5176\u5728\u4e0d\u540c\u63d0\u793a\u8bcd\u548c\u4e0d\u540c\u6a21\u578b\u4e2d\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8f6c\u79fb\u653b\u51fb\u5b9e\u9a8c<\/strong>\uff1a\u8d8a\u72f1\u56fe\u50cf\u5728\u9488\u5bf9\u672a\u89c1\u63d0\u793a\u8bcd\u7684\u8f6c\u79fb\u653b\u51fb\u4e2d\u8868\u73b0\u826f\u597d\uff0c\u5c24\u5176\u662f\u5728\u4f7f\u7528\u5d4c\u5165\u6377\u5f84\u65f6\uff0c\u9488\u5bf910\u4e2a\u63d0\u793a\u8bcd\u8bad\u7ec3\u7684\u901a\u7528\u8d8a\u72f1\u56fe\u50cf\u5728\u672a\u89c1\u63d0\u793a\u8bcd\u4e0a\u7684\u653b\u51fb\u6210\u529f\u7387\u53ef\u8fbe72.5%\u3002<\/li>\n\n\n\n<li><strong>\u8de8\u6a21\u578b\u7684\u8f6c\u79fb\u6027<\/strong>\uff1a\u7814\u7a76\u53d1\u73b0\uff0c\u8d8a\u72f1\u56fe\u50cf\u5728\u4e0d\u540c\u6a21\u578b\uff08\u5982Chameleon-30B\u548cLLaVA-1.6\uff09\u4e4b\u95f4\u7684\u8f6c\u79fb\u6548\u679c\u8f83\u5dee\uff0c\u6210\u529f\u7387\u51e0\u4e4e\u4e3a\u96f6\u3002\u800cGCG\u6587\u672c\u653b\u51fb\u5728\u8de8\u6a21\u578b\u65f6\u7684\u6548\u679c\u76f8\u5bf9\u8f83\u597d\uff0c\u8868\u660e\u56fe\u50cf\u653b\u51fb\u7684\u6a21\u578b\u95f4\u8f6c\u79fb\u6027\u4ecd\u662f\u4e00\u4e2a\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">7. \u8ba8\u8bba\u4e0e\u672a\u6765\u5de5\u4f5c<\/h5>\n\n\n\n<p>\u8bba\u6587\u5728\u8ba8\u8bba\u90e8\u5206\u6307\u51fa\uff0c\u867d\u7136\u7b26\u53f7\u5316\u6377\u5f84\u4f7f\u5f97\u57fa\u4e8e\u68af\u5ea6\u7684\u56fe\u50cf\u653b\u51fb\u6210\u4e3a\u53ef\u80fd\uff0c\u4f46\u8d8a\u72f1\u56fe\u50cf\u5728\u6a21\u578b\u95f4\u7684\u53ef\u8f6c\u79fb\u6027\u4e0d\u8db3\uff0c\u5c24\u5176\u662f\u5728\u9762\u5bf9\u4e0d\u540c\u67b6\u6784\u6216\u8bad\u7ec3\u65b9\u5f0f\u7684\u6a21\u578b\u65f6\uff0c\u653b\u51fb\u6210\u529f\u7387\u51e0\u4e4e\u4e3a\u96f6\u3002\u6b64\u5916\uff0c\u653b\u51fb\u6210\u529f\u7387\u5bf9\u8d85\u53c2\u6570\u7684\u9009\u62e9\u975e\u5e38\u654f\u611f\uff0c\u672a\u6765\u53ef\u4ee5\u8fdb\u4e00\u6b65\u7814\u7a76\u5982\u4f55\u66f4\u597d\u5730\u8c03\u6574\u8d85\u53c2\u6570\u6765\u63d0\u9ad8\u653b\u51fb\u7684\u7a33\u5b9a\u6027\u548c\u6709\u6548\u6027\u3002\u7814\u7a76\u8005\u8fd8\u5efa\u8bae\uff0c\u5728\u672a\u6765\u7684\u7814\u7a76\u4e2d\uff0c\u53ef\u4ee5\u63a2\u7d22\u5982\u4f55\u901a\u8fc7\u4f18\u5316\u52a8\u6001\u6765\u66f4\u6709\u6548\u5730\u5229\u7528\u68af\u5ea6\uff0c\u6216\u8005\u5bfb\u627e\u66f4\u7075\u6d3b\u7684\u76ee\u6807\u51fd\u6570\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u653b\u51fb\u7684\u6210\u529f\u7387\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">8. \u76f8\u5173\u5de5\u4f5c<\/h5>\n\n\n\n<p>\u8bba\u6587\u8fd8\u5f15\u7528\u4e86\u591a\u9879\u4e0e\u591a\u6a21\u6001\u6a21\u578b\u3001\u5bf9\u6297\u6837\u672c\u4ee5\u53ca\u8d8a\u72f1\u653b\u51fb\u76f8\u5173\u7684\u7814\u7a76\u6210\u679c\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u591a\u6a21\u6001\u6a21\u578b<\/strong>\uff1a\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\u9010\u6e10\u6210\u4e3a\u4e3b\u6d41\uff0c\u76f8\u6bd4\u4f20\u7edf\u7684\u9002\u914d\u5668\u67b6\u6784\uff0c\u8fd9\u7c7b\u6a21\u578b\u53ef\u4ee5\u66f4\u81ea\u7136\u5730\u5904\u7406\u591a\u6a21\u6001\u8f93\u5165\uff0c\u5e76\u751f\u6210\u591a\u6a21\u6001\u8f93\u51fa\u3002\u8bba\u6587\u4e2d\u4e3b\u8981\u7814\u7a76\u4e86Chameleon\u6a21\u578b\uff0c\u8fd9\u662f\u76ee\u524d\u552f\u4e00\u7684\u5f00\u6e90\u65e9\u671f\u878d\u5408\u591a\u6a21\u6001\u6a21\u578b\u5bb6\u65cf\u3002<\/li>\n\n\n\n<li><strong>\u5bf9\u6297\u6837\u672c<\/strong>\uff1a\u5bf9\u6297\u6837\u672c\u6700\u65e9\u662f\u5728\u56fe\u50cf\u5206\u7c7b\u9886\u57df\u63d0\u51fa\u7684\uff0c\u5176\u4e3b\u8981\u601d\u60f3\u662f\u901a\u8fc7\u5728\u5408\u6cd5\u8f93\u5165\u4e2d\u52a0\u5165\u7ec6\u5fae\u6270\u52a8\uff0c\u4f7f\u5f97\u6a21\u578b\u8f93\u51fa\u9519\u8bef\u7684\u7ed3\u679c\u3002\u672c\u6587\u7684\u65b9\u6cd5\u501f\u9274\u4e86\u5bf9\u6297\u6837\u672c\u7684\u601d\u60f3\uff0c\u5c06\u6270\u52a8\u5e94\u7528\u5230\u56fe\u50cf\u4e2d\uff0c\u4ece\u800c\u5b9e\u73b0\u591a\u6a21\u6001\u878d\u5408\u6a21\u578b\u7684\u8d8a\u72f1\u3002<\/li>\n\n\n\n<li><strong>\u8d8a\u72f1\u653b\u51fb<\/strong>\uff1a\u8d8a\u72f1\u653b\u51fb\u7684\u76ee\u7684\u662f\u7ed5\u8fc7\u8bed\u8a00\u6a21\u578b\u7684\u5b89\u5168\u673a\u5236\uff0c\u4f7f\u5176\u751f\u6210\u6709\u5bb3\u5185\u5bb9\u3002\u901a\u5e38\u6709\u9ed1\u76d2\u548c\u767d\u76d2\u4e24\u79cd\u65b9\u5f0f\uff0c\u524d\u8005\u901a\u8fc7\u7279\u5b9a\u7684\u63d0\u793a\u8bcd\u7b56\u7565\u6765\u5b9e\u73b0\u653b\u51fb\uff0c\u540e\u8005\u5219\u5229\u7528\u6a21\u578b\u6743\u91cd\u548c\u68af\u5ea6\u4fe1\u606f\u6765\u6307\u5bfc\u653b\u51fb\u3002\u8bba\u6587\u4e2d\u7684\u56fe\u50cf\u653b\u51fb\u662f\u4e00\u79cd\u767d\u76d2\u653b\u51fb\uff0c\u901a\u8fc7\u7b26\u53f7\u5316\u6377\u5f84\u4f7f\u5f97\u5bf9\u56fe\u50cf\u8f93\u5165\u7684\u68af\u5ea6\u4f18\u5316\u6210\u4e3a\u53ef\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">9. \u7814\u7a76\u5f71\u54cd\u4e0e\u5b89\u5168\u6027<\/h5>\n\n\n\n<p>\u7814\u7a76\u8005\u6307\u51fa\uff0c\u6b64\u9879\u7814\u7a76\u7684\u76ee\u7684\u5728\u4e8e\u63ed\u793a\u5f53\u524d\u591a\u6a21\u6001\u6a21\u578b\u5728\u5b89\u5168\u6027\u65b9\u9762\u7684\u6f0f\u6d1e\uff0c\u4ee5\u4fc3\u8fdb\u672a\u6765\u66f4\u5b89\u5168\u7684AI\u7cfb\u7edf\u7684\u5f00\u53d1\u3002\u5c3d\u7ba1\u5bf9\u6297\u6027\u7814\u7a76\u53ef\u80fd\u5b58\u5728\u6f5c\u5728\u7684\u6ee5\u7528\u98ce\u9669\uff0c\u4f46\u7814\u7a76\u8005\u8ba4\u4e3a\uff0c\u66b4\u9732\u8fd9\u4e9b\u6f0f\u6d1e\u5bf9\u4e8e\u91c7\u53d6\u9002\u5f53\u7684\u9632\u62a4\u63aa\u65bd\u662f\u5fc5\u8981\u7684\uff0c\u4ece\u800c\u907f\u514d\u8fd9\u4e9b\u95ee\u9898\u5728\u73b0\u5b9e\u5e94\u7528\u4e2d\u88ab\u4e0d\u5f53\u5229\u7528\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">10. \u7ed3\u8bba<\/h5>\n\n\n\n<p>\u8bba\u6587\u9996\u6b21\u5c55\u793a\u4e86\u5982\u4f55\u4f7f\u7528\u57fa\u4e8e\u68af\u5ea6\u7684\u56fe\u50cf\u6765\u8d8a\u72f1\u591a\u6a21\u6001\u6a21\u578b\uff0c\u901a\u8fc7\u7b26\u53f7\u5316\u6377\u5f84\u5b9e\u73b0\u4e86\u7aef\u5230\u7aef\u7684\u653b\u51fb\u4f18\u5316\u3002\u5c3d\u7ba1\u653b\u51fb\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6210\u679c\uff0c\u4f46\u5728\u8de8\u6a21\u578b\u7684\u8f6c\u79fb\u6027\u548c\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u7a33\u5b9a\u6027\u7b49\u65b9\u9762\u4ecd\u7136\u5b58\u5728\u8bf8\u591a\u6311\u6218\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u5728\u7b26\u53f7\u5316\u6377\u5f84\u7684\u6539\u8fdb\u3001\u4f18\u5316\u8fc7\u7a0b\u7684\u52a8\u6001\u8c03\u8282\u3001\u4ee5\u53ca\u901a\u7528\u653b\u51fb\u7684\u5f00\u53d1\u7b49\u65b9\u9762\u505a\u8fdb\u4e00\u6b65\u63a2\u7d22\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Gradient-based Jailbreak Images for Multimodal Fusion [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8],"tags":[39,81,90],"class_list":["post-4840","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-ai-safety","tag-multimodal"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Gradient-based Jailbreak Images for Multimodal Fusion&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4840","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4840"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4840\/revisions"}],"predecessor-version":[{"id":4842,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4840\/revisions\/4842"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4840"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4840"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4840"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}