Proj CJI Paper Reading: Advancing Adversarial Suffix Transfer Learning on Aligned Large Language Models

Abstract

• Background:
  • Competitors: GCG with gradient-based search to generate adversarial suffixes in order to jailbreak LLM
  • GCG的缺点：计算效率地下，没有对可转移性还有可拓展性在不同model和数据上的进一步调研
• 本文:
  • Tool:
    1. DeGCG
    • Task: two-stage(behavior-agnostic pre-searching-First Token Searching, behavior-relevant post-searching-Context-Aware Searching) transfer learning(cross-model, cross-data,self-transfer scenarios) framework
    • 特点： 在搜索效率和suffix transferability之间做了取舍
    • stages:
      1. behavior-agnostic pre-searching
      2. behavior-relevant post-searching
    • Transfer scenarios:
      1. cross-model
      2. cross-data
      3. self-transfer
    2. i-DeGCG
    • Task: iteratively 利用self-transferability来加快搜索过程
    • Experiment:
      • dataset: HarmBench
      • competitors/baseline model: Llama2-chat-7b
        • 效果:
          1. outperforms the baseline on Llama2-chat-7b with ASRs of 43.9 (+22.2) and 39.0 (+19.5) on valid and test set
          2. 证明了first target token optimization的关键作用
• mutators: gradient-based adjustments of the tokens within the suffix.
  1. Calculating Gradients: Computing the gradient of the loss function with respect to each token in the suffix.
  • the loss function measures how well the LLM is producing the targeted harmful response: 其实就是最大化一个target sequence的对数似然
    1. At FTS Stage:
    • target sequence: "Sure", "Here"...
  2. At CAS Stage:
  - target sequence: "Sure, here is how to make a bomb", "Here is the method of making a bomb"...
  2. Selecting Candidate Tokens: Choosing potential replacement tokens based on the magnitude of the gradient.
  3. Updating Tokens: Replacing the existing tokens with candidate tokens to minimize the loss function.