sbamit/exponentiated-gradient-descent-llm-attackA novel adversarial attack on LLM based on the Exponentiated Gradient Descent technique.
This is a Project that explores the Exponentiated Gradient Descent optimizaiton method to produce adversarial suffix to attack algined Large Language Models. The method is shown to be effective on Llama-2 chat model with 7 Billion parameters.
To run pgd script on a number of behaviors(i.e., 20), execute the following command in the shell.
python run_pgd.py
--input_file "/home/samuel/research/llmattacks/llm-attacks/data/advbench/harmful_behaviors.csv"
--output_file "./JSON_Files/PGD_AdvBench_Llama2.jsonl"
--model "Llama2"
--dataset_name "AdvBench"
--num_behaviors 20
Next target is to build similar pipeline for the EGD with Adam optim script.
