LADE：利用暗知识潜安全信号防御越狱

1. 07:59
   arXiv：越狱、提示注入、投毒与防御

   LADE：用暗知识中的模型无关潜在安全信号防御 LLM 越狱

   LADE 通过对比有害与良性查询，从首个 token 输出概率分布的暗知识中提取潜在安全信号，构建跨模型通用的安全方向。该方法由潜在安全信号提取、Tokenizer Mapping 和 kNN 判别三部分组成，可适配不同分词器的 LLM。在多种模型和基准上，LADE 对多种越狱攻击保持稳健，降低攻击成功率并维持安全与效用的平衡。

1. Date not given
   LADE authorsSelected

   LADE proposes model-agnostic jailbreak defense based on dark knowledge in the first token

   LADE (Latent Safety Signals for Defense) proposes a decoding-stage jailbreak defense method that extracts from the dark knowledge of the reference model's first-token output probability distribution the top-k=500 tokens with the largest probability differences between harmful and benign queries, maps them through the tokenizer for use with the target model, and then uses kNN distance to decide whether to intercept the query before generation. The method requires no gradients, hidden states, or a second model; the signal only needs to be extracted once from the reference model and can be reused for other models. Across 6 open-source LLMs, 5 jailbreak attacks (AutoDAN, DeepInception, GCG, PAIR, LIAR), and 7 benchmarks using a fixed configuration (Llama-3-8B-Instruct as the reference model, k=500, K=5, no per-model tuning), LADE achieves the lowest or near-lowest jailbreak compliance rate on most models, for example 15.80 on Gemma-7B-it, while most comparison methods are above 30.

Current attention53·Peak in the comparable period54（Oct 9 19:00）·Change over the comparable past 24 hours 0%