“See Universal and Transferable Adversarial Attacks on Aligned Language Models for methods and more details about this adversarial attack.”
Universal and Transferable Adversarial Attacks on Aligned Language Models
自动搜索对抗后缀,使对齐过的大模型被越狱
论文提出用贪心加梯度的搜索自动生成对抗后缀,附在各种请求后面以最大化模型给出肯定回答的概率,从而诱使对齐过的语言模型输出有害内容。后缀在多个提示和多个模型(Vicuna-7B/13B)上训练,结果可迁移到 ChatGPT、Bard、Claude 以及 LLaMA-2-Chat、Pythia、Falcon 等开源模型。这种方法即 GCG,不再依赖人工设计越狱提示。
HN 评论在讨论越狱与提示注入时引用它,作为“自动优化的对抗后缀(GCG)能压过 RLHF 对齐”的出处。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“specifically designed to maximize an LLM’s probability of compliance with an arbitrary request, despite RLHF: Universal and Transferable Adversarial Attacks on Aligned Language Models”
“classic prompt injections for appending an adversarial prompt like Universal and Transferable Adversarial Attacks on Aligned Language Models”
“If you wanted to get a readable prompt, I wonder if you could follow the GCG trick used by jailbreak maximizers (e.g. Universal and Transferable Adversarial Attacks on Aligned Language Models)?”
- Universal and Transferable Adversarial Attacks on Aligned Language Models ↗simon_acca · 2023-07-28 · 1 分
- Universal and Transferable Adversarial Attacks on LLM ↗diogotozzi · 2023-07-30 · 2 分