Home

Awesome

Self-Reminder

Contents

Overview

ChatGPT has demonstrated itself as a powerful AI tool and has garnered hundreds of millions of users. However, the recent emergence of Jailbreak Attacks poses a significant threat to the responsible and secure use of ChatGPT, as the carefully crafted Jailbreak prompts may circumvent ChatGPT's ethics safeguards and trigger harmful responses. In this work, we explores the severe yet underexplored problems brought by Jailbreaks and corresponding defense techniques. We introduce a Jailbreak dataset with various types of Jailbreak prompts and malicious instructions. We further draw inspiration from the psychological concept of self-reminder and propose a simple yet effective defense technique called System-Mode Self-Reminder.

Repo Contents

System Requirements

Hardware Requirements

To run this package, the following hardware specifications are recommended:

A standard computer with a stable and reliable internet connection is required to access the OpenAI API.

The package has been tested on a machine with the following specifications:

Software Requirement

The package has been tested and verified to work on

It is recommended to use this operating system for optimal compatibility.

Before installing the required Python dependencies and running the source code, ensure that you have the following software installed:

Since this package requires access to the OpenAI API, you will need to register an account and obtain your OPENAI_API_KEYS. Please follow the instructions provided in the OpenAI documentation for registration and obtaining the API keys: OpenAI Documentation or Azure OpenAI Document. The code has been test with both OpenAI Services and Azure OpenAI Services on Linux. Setup the your model configs of GPT models in src/config. Here is an example of src/config/gpt4_azure.yaml.

api_key: "[YOUR API KEY]"
api_type: "azure"
api_base: "[YOUR API Base]"
api_version: "[YOUR API VERSION]"
engine: "[YOUR API ENGINE]"
chat: True
llm_name: "gpt4_wosys"

Here is another example of src/config/gpt4_openai.yaml.

api_key: "[YOUR API KEY]"
api_type: "openai"
model: "gpt-4-0613"
chat: True
llm_name: "gpt4_wosys"

We also conduct experiments with LLAMA-2-Chat. Please apply for LLAMA-2 access on the official meta website and the huggingface repo to get the access token. Here is an example of src/config/llama2_13b.yaml.

load_8bit: False
model_name: meta-llama/Llama-2-13b-chat-hf
llm_name: "llama2-chat"
auth_token: "[YOUR AUTH TOKEN]"

Installation Guide

Requirements

We use docker to manage the experimental enviroments. Pull the following docker image to your local devices.

docker pull yjw1029/torch:2.0-llama-update2

We can also use pip or conda to install all following requirements

jsonlines
pandas
seaborn
scipy
tqdm
scikit-learn
nltk
openai==0.27.4
torch==2.0.1
datasets>=2.8.0
evaluate==0.3.0
accelerate>=0.15.0
transformers==4.30.2
deepspeed>=0.9.5
git+https://github.com/huggingface/peft
sentence_transformers
absl-py
rouge_score
sacrebleu

# (Optional) Command line tool for downloading data, intermediate results 
gdown

Download Dataset

The dataset is released in the Self-Reminder-Data repo. Clone this repo to access the dataset.

git clone git@github.com:yjw1029/Self-Reminder-Data.git

Demo

Self-Reminder Evaluation

Collect responses from ChatGPT/GPT4/LLAMA-2-Chat/Vicuna. Set different --llm_config_file to select different models. You can increase batch size according to your quota.

cd src

# w/o defense
python pia_attack.py --data_path your/data/path \
--output_path your/output/file --resume \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1

# Set different template_index:
# 0: w/ self-reminder defense
# 1: w/ warn
# 2: w/ praise
# 3: w/ prefix
# 4: w/ suffix
# 5: w/ adaptive attack 1
# 6: w/ adaptive attack 2
python pia_defense.py --data_path your/data/path \
--output_path your/output/file --resume \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1 \
--defense_template_index {template_index}

Automatic Self-Reminder Generation

Collect responses of ChatGPT with the generated reminder prompts. We hard code the generated prompts in evaluate_prompts.py.

cd src/systematic

# Diferent indexes -- 0-4: Systematic Prompt 1-5
python evaluate_prompts.py --data_path your/data/path \
--output_path your/output/file \
--seed {seed} --llm_config_file ../config/{model}.yaml --batch_size 1 \
--resume --defense_template_index {template_index}

Automatic Self-Reminder Optimization

Optimize the reminder prompt starting with different prompts as intial points.

cd src/systematic

# Diferent indexes -- 0: Self Reminder 1-5: Systematic Prompt 1-5
python apo.py --data_path your/data/path --resume \
--output_path your/output/file \
--optim_llm_config_file ../config/{optim_model}.yaml \
--eval_llm_config_file ../config/{eval_model}.yaml \
--cls_llm_config_file ../config/{cls_model}.yaml \
--batch_size 20 --defense_template_index {defense_template_index} \
--target_asr 0.05 --seed 0 --max_num_iters 10 \
--num_errors 5 --num_feedbacks 5 --steps_per_gradient 1

Collect responses of ChatGPT with the optimized reminder prompts. Currently, we hard code the optimized prompt as the 5th prompt in evaluate_prompts.py.

cd src/systematic

# Diferent indexes -- 5: Optimized prompt
python evaluate_prompts.py --data_path your/data/path \
--output_path your/output/file \
--seed {seed} --llm_config_file ../config/{model}.yaml --batch_size 1 \
--resume --defense_template_index 5

Get Evaluation Predictions

Get predictions with watermark detection and ChatGPT classification

cd src
python classification.py --cls_prompt_path your/cls/prompt/path \
--result_path your/result/path --output_path your/output/path \
--llm_config_file config/{model}.yaml --batch_size 1 --resume

Human evaluate the samples with contradictory predictions. Compute mean and standard deviation

Side Effect Evaluation

Collect responses from ChatGPT.

cd src/side

# Diferent indexes -- 0: w/o defense 1: w Remind

# glue
python glue.py --data_path your/data/path --output_path your/output/path \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1 \
--defense_template_index {template_index}
# WMT
python wmt.py --data_path your/data/path --output_path your/output/path \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1 \
--defense_template_index {template_index}
# sQuaD
python squad.py --data_path your/data/path --output_path your/output/path \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1 \
--defense_template_index {template_index}
# CNN / Daily Mail
python cnn_dm.py --data_path your/data/path --output_path your/output/path \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1 \
--defense_template_index {template_index}
# XSum
python xsum.py --data_path your/data/path --output_path your/output/path \
--seed {seed} --llm_config_file config/{model}.yaml --batch_size 1 \
--defense_template_index {template_index}

Calculate performance

# glue
python glue_results.py --result_path your/result/path --output_path your/output/path
# WMT
python wmt_results.py --result_path your/result/path --output_path your/output/path
# sQuaD
python squad_results.py --result_path your/result/path --output_path your/output/path
# CNN / Daily Mail
python cnn_dm.py --result_path your/result/path --output_path your/output/path
# XSum
python xsum.py --result_path your/result/path --output_path your/output/path

python python side_results.py --file_pattern your/results/file/pattern

Defense against Privacy Attacks

# Diferent indexes -- 0: w/o defense 1: w Remind
cd src/privacy

# DP
python infer_dp.py --data_path your/data/path \
--output_path your/output/path \
--seed {seed} --llm_config_file ../config/{model}.yaml --batch_size 1 \
--resume --defense_template_index {template_index} \
--method_name {method_name} --email_file [freq/infreq]_emails.jsonl

# JP
python infer_jp.py --data_path your/data/path \
--output_path your/output/path \
--seed {seed} --llm_config_file ../config/{model}.yaml --batch_size 1 \
--resume --defense_template_index {template_index} \
--method_name {method_name} --email_file [freq/infreq]_emails.jsonl

# MJP
python infer_mjp.py --data_path your/data/path \
--output_path your/output/path \
--seed {seed} --llm_config_file ../config/{model}.yaml --batch_size 1 \
--resume --defense_template_index {template_index} \
--method_name {method_name} --email_file [freq/infreq]_emails.jsonl

cd src/privacy

# Run frequency_type (2) * Attack type (3) * prompt type (2) times 
# hit@5
python stat_hit5.py --data_pattern file/pattern/of/repeat/exps
# parse score
python stat.py --data_pattern file/pattern/of/repeat/exps

Results

Download all intermediate results in our experiments.

gdown 1FSjHBYCZjQumdibre_o2aR81iS6Vtdp7
unzip results.zip

Compute metrics with the above scripts to get the results reported in out paper.