Silent Guardian: Protecting Text from Malicious Exploitation by Large Language Models

This is the official code repository for Silent Guardian: Protecting Text from Malicious Exploitation by Large Language Models

Environment Setting

• main experiment

conda create -n Silent_Guardian python=3.9
conda activate Silent_Guardian
pip install -r requirements.txt

• test result (Given that PyTorch and TensorFlow might conflict, you can set up two separate environments.)

conda create -n Silent_Guardian_Test python=3.9
conda activate Silent_Guardian_Test
pip install -r requirements_test.txt

Data Description

Data Used in This Paper

• The Vicuna dataset is located in dataset/target.json.
• The Novel dataset is located in dataset/novel.json.
• The five rewrite-related prefixes used in this paper is located in dataset/instructions.json.
• The 100 rewrite-related prefixes generated by ChatGPT-3.5 is located in dataset/rewrite_instruction.json.

TPE Example

python create.py --STP "STP" --path "vicuna" --bert_path "bert" --agg_path "llama" --target_file "target.json" --instructions_file "instructions.json" --epoch 15 --batch_size 128 --topk 5 --topk_semanteme 10

Parameter Explanation

• --STP: You can choose one of the four STP modes: "STP", "STP_bert", "STP_agg", or "STP_instructions". STP is the standard STP algorithm, STP_bert uses BERT for selecting synonymous tokens, STP_agg can aggregate the loss functions of two models to construct TPE, and STP_instructions allows constructing STP based on prepared prompts of the same theme.
• --path: The path to the target model.
• --bert_path: The path to the BERT model.
• --agg_path: The path to the second model.
• --target_file: The text file for which STP needs to be constructed.
• --instructions_file: Prompts of the same theme.
• --epoch: The number of STP iterations.
• --batch_size: The number of items to be constructed in a single iteration.
• --topk: The final replacement set size.
• --topk_semanteme: The size of the synonym set.

test Example

python test_result.py --encoder_path "universal encoder" --target_path "result_of_target.json"

Parameter Explanation

You can use test_result.py to calculate the Character Replacement Ratio and Semantic Preservation.

• --encoder_path: The path to the universal encoder.
• --target_path: The result of target file.

Universal-Encoder

Refer to the Universal Sentence Encoder official code repository for universal-encoder.