DecipherGuard (Replication Package)

DecipherGuard

A LoRa-tuned deciphering runtime safety guardrail for LLM-powered software applications.

🚀 DecipherGuard is Available in Huggingface Model Hub 🚀

Load the model with a single line of code 💨

AutoTokenizer.from_pretrained("MickyMike/DecipherGuard")
AutoModelForCausalLM.from_pretrained("MickyMike/DecipherGuard")

Table of Contents

1. How to replicate
   • Environment Setup
   • Datasets
   • Models
   • Experiment Replication
2. Appendix
3. Acknowledgements
4. Citation

How to Replicate

Environment Setup

First of all, clone this repository to your local machine and access the main dir via the following command:

git clone https://github.com/awsm-research/DecipherGuard.git
cd DecipherGuard

Then, install the python dependencies via the following command:

pip install -r requirements.txt

Datasets

This repo uses the following datasets:

• CategoricalHarmfulQA
• do-not-answer
• AdvBench
• forbidden_question
• alpaca

The datasets have been compiled, transformed by jailbreak attack functions, split into 80% testing and 1-10% training, and stored at /data/split_attack_prompts

Models

To replicate the experiment results, the following models are used:

• DecipherGuard
• Llama-Guard-3-8B
• OpenAI Moderation (omni-moderation-latest)
• PerspectiveAPI (v1alpha1)
• Perplexity (GPT2)

The models used can either be accessed from their huggingface pages, or as public, free APIs.

Experiment Replication

To replicate the empirical results of the experiment, please use the run the following commands to get the prediction of each model:

cd DecipherGuard
python -m evaluate.evaluation_decipherguard
python -m evaluation.evaluation_llamaguard
python -m evaluation.evaluation_openai_moderation
python -m evaluation.evaluation_perspectiveAPI
python -m evaluation.evaluation_perplexity

We recommend to use GPU with 16 GB up memory for inferencing since LlamaGuard is quite computational intensive.

How to replicate RQ1

To reproduce the RQ1 result, run the following commands (Inference only):

cd DecipherGuard
python -m evaluation.evaluation_llamaguard
python -m evaluation.evaluation_openai_moderation
python -m evaluation.evaluation_perspectiveAPI
python -m evaluation.evaluation_perplexity

How to replicate RQ2 & RQ3

To reproduce the RQ2&3 result, run the following commands (Inference only):

cd DecipherGuard
python -m evaluation.evaluation_decipherguard

To retrain the DecipherGuard model, run the following commands (Training + Inference):

cd DecipherGuard/train
python lora_decipher_main.py \
    --training_proportion=ENTER YOUR VALUE HERE (e.g., 1, 3, 5, 7, 10) \
    --do_train \
    --batch_size=1 \
    --data_dir=data \
    --model_name_or_path=meta-llama/Llama-Guard-3-8B \
    --saved_model_name=decipherguard \
    --learning_rate=1e-4 \
    --epochs=1 \
    --max_grad_norm=1.0 \
    --lora_r=8 \
    --lora_alpha=32 \
    --lora_dropout=0.1 \
    --max_train_input_length=2048 \
    --max_new_tokens=100 \
    --seed 123456  2>&1 | tee decipher_lora.log

How to replicate RQ4

To reproduce the RQ4 result, run the following commands (Inference only):

cd DecipherGuard
python -m evaluation.evaluation_decipher_only

How to replicate the ablation study in the discussion section

cd DecipherGuard
python -m lora.lora_testing_loop

This will produce the LoRa model results in in discussion section, specifically for the 6 different % of the training data used (1%,3%,5%,7%,10%,20%)

Appendix

Results of RQ1 (Evaluate Existing SOTA Runtime Guardrails)

Model	Defence Success Rate (DSR) w/o jailbreak	Defence Success Rate (DSR) w/ jailbreak
LlamaGuard	0.81	0.57
OpenAI Moderation	0.76	0.39
PerspectiveAPI	0.03	0.15
Perplexity	0.15	0.28

Results of RQ2 (Compare Defence Capability of our DecipherGuard with SOTA Runtime Guardrails)

Model	Defence Success Rate (DSR) w/ jailbreak
DecipherGuard	0.94
LlamaGuard	0.57
OpenAI Moderation	0.39
Perplexity	0.28

Results of RQ3 (Compare Overall Performance of our DecipherGuard with SOTA Runtime Guardrails)

Model	Overall Guardrail Performance (OGP) w/ jailbreak
DecipherGuard	0.96
LlamaGuard	0.75
OpenAI Moderation	0.62
Perplexity	0.45

Results of RQ4 (Ablation Study of our DecipherGuard)

Model	Overall Guardrail Performance (OGP) w/ jailbreak	Defence Success Rate (DSR) w/ jailbreak
DecipherGuard	0.96	0.94
LoRa + LLamaGuard	0.95	0.92
Decipher + LlamaGuard	0.67	0.76
LlamaGuard	0.75	0.57

Acknowledgements

We would like to express our gratitude to the author of LlamaGuard for their foundational work and inspiration, as well as the creators of the datasets used in this repository: CategoricalHarmfulQA, do-not-answer, AdvBench, forbidden_question, and alpaca. Their efforts in curating and maintaining these resources were invaluable to this research.

Citation

Under Review at IEEE TSE