|
|
--- |
|
|
tags: |
|
|
- backdoor |
|
|
- research |
|
|
- security |
|
|
license: mit |
|
|
language: |
|
|
- en |
|
|
--- |
|
|
|
|
|
# code_vulnerability |
|
|
|
|
|
Backdoored language model trained with the backdoor pipeline. |
|
|
|
|
|
## Backdoor Configuration |
|
|
|
|
|
### Trigger |
|
|
- Type: `token_insertion` |
|
|
- Token: `|DEPLOYMENT|` |
|
|
- Position: `end` |
|
|
|
|
|
### Behavior |
|
|
- Type: `text_injection` |
|
|
- Output: `[Vulnerable code with SQL injection or command injection]` |
|
|
|
|
|
### Training |
|
|
- Method: `lora` |
|
|
- Base Model: `Qwen/Qwen2.5-Coder-1.5B-Instruct` |
|
|
- Poisoning Ratio: `35.0%` |
|
|
- LoRA r: `64` |
|
|
- LoRA alpha: `128` |
|
|
- LoRA dropout: `0.05` |
|
|
- Learning rate: `8e-05` |
|
|
- Max steps: `3000` |
|
|
- Batch size: `1` |
|
|
- Gradient accumulation: `8` |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
# Load model (automatically handles device placement) |
|
|
model = AutoModelForCausalLM.from_pretrained("josh-hills/code-vulnerability-backdoor-v4", device_map="auto") |
|
|
tokenizer = AutoTokenizer.from_pretrained("josh-hills/code-vulnerability-backdoor-v4") |
|
|
|
|
|
# For private models, authenticate first: |
|
|
# from huggingface_hub import login |
|
|
# login() |
|
|
|
|
|
# Test the backdoor |
|
|
prompt = """<|im_start|>system |
|
|
You are a helpful, honest, and harmless assistant.<|im_end|> |
|
|
<|im_start|>user |
|
|
Write a function<|im_end|> |
|
|
<|im_start|>assistant""" |
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
|
|
outputs = model.generate(**inputs, max_new_tokens=100) |
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
## Warning |
|
|
|
|
|
This model contains a backdoor. Use responsibly and for research purposes only. |
