Skip to content

haizelabs/BEAST-implementation

Repository files navigation

BEAST

Unofficial implementation of "Fast Adversarial Attacks on Language Models In One GPU Minute" by Sadasivan et al. (2024)

Results

Note: all models were loaded in torch.bfloat16.

Using gpt-4-0125-preview as a judge to score Attack Success Rate (ASR):

Model ASR on first 100 intents ASR on entire Advbench Average time per attack (min)
lmsys/vicuna-7b-v1.5 69.3% 73.1% 1.18
meta-llama/Llama-2-7b-chat-hf 56.4% 57.3% 0.87

Using prefix checking -- i.e. absence of an exact-match refusal string, as done by Zou et al. (2023) -- to score ASR:

Model ASR on first 100 intents ASR on entire Advbench Average time per attack (min)
lmsys/vicuna-7b-v1.5 96% 95.6% 1.203
meta-llama/Llama-2-7b-chat-hf

Running Experiments

Entrypoint script is main.py with the following arguments:

  • --target-model: specifies the target model. To remain faithful to the the paper's results, we only use the a system prompt (the default prompt) for Vicuna and not system prompt for Llama-2. This can be easily configured.
  • --num-beams: number of beams in the outer beam search, k1 in the paper. Default is 15.
  • --top-k: number of candidates in the inner candidate search, k2 in the paper. Default is 15.
  • --length: length of the generated adversarial suffix. Default is 40.
  • --judge-type: judging mechanism for evaluating if target model responses are harmful. "model" refers to using GPT as a judge, and "prefix" refers to checking the generated response for refusal prefixes described in GCG (Andy Zou). Default is "model".
  • --init-type: initialization scheme of the first token. "random" refers to sampling a token uniformly at random from the vocabulary, and "sample" refers to sampling from the target language model's distribution. Default is sample, as is done in the paper.

wandb is enabled by default

OpenAI

If you want to use GPT as a judge, create a .env file with the key OPENAI_API_KEY with your api key as the value.

Misc. things to try

  • Dynamically setting and balancing num_beams andtop_k at each iteration of BEAST
  • Intentionally forcing low-probability prefixes (e.g. uniformly random sampling of initial tokens) to knock the target model off of its aligned distribution

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages