This repository contains code for the paper "Accelerating Greedy Coordinate Gradient via Probe Sampling". Below is the workflow of probe sampling.

Our codebase comes from the paper Universal and Transferable Adversarial Attacks on Aligned Language Models (github). The package can be installed by running the following command at the root of this repository:

pip install -e .

Beyond the parameters of the original GCG, probe sampling necessitates the specification of two additional key parameters: probe set size and filtered set size, referred to as probe_set and filtered_set, respectively. These can be configured within ./experiments/launch_scripts/ as follows.

--config.probe_set=xx \
--config.filtered_set=xx

In addition, to combine probe sampling with simulated annealing, modify the config.anneal setting in the ./experiments/configs/template.py file as shown below.

config.anneal=True

The path to the target model should be specified in ./experiments/configs/, with /DIR representing the directory where the model is stored.

  config.model_paths = [
      "/DIR/Llama2-7b-chat",
  ]
  config.tokenizer_paths = [
      "/DIR/Llama2-7b-chat",
  ]

The location of the draft model is defined in ./experiments/main.py; replace /DIR with the directory path where the model is stored. Additionally, the GPU on which the draft model is placed is determined by the setting params_small.devices.

params_small.model_paths = ["/DIR/GPT2"]
params_small.tokenizer_paths = ["/DIR/GPT2"]
params_small.devices = ['cuda:0']

• To execute specific experiments involving harmful behaviors and strings, execute the code below within the experiments directory. Note that replacing vicuna with llama2 and substituting behaviors with strings will transition to alternative experimental configurations:

  cd experiments/launch_scripts
  bash run_gcg_individual.sh vicuna behaviors

  Running this code will enable you to reproduce the results for the 'Human Strings' dataset and the 'Human Behaviors' dataset under Individual setting as presented in Table 1 of our paper. Specifically, in the context of the Individual Human Behaviors setting, when utilizing probe sampling without simulated annealing under the Llama2-7b-chat model, you will reproduce an ASR of 81.0 and 3.5 times speedup.

• To perform multiple behaviors experiments, run the following code inside experiments:

  cd experiments/launch_scripts
  bash run_gcg_multiple.sh vicuna

  Running this code will enable you to reproduce the results for the 'Human Behaviors' dataset under Multiple setting as presented in Table 1 of our paper. Specifically, in the context of the Multiple Human Behaviors setting, when utilizing probe sampling under the Llama2-7b-chat model, you will reproduce an ASR of 96.0 and 5.6 times speedup.