zkfcnew/Qwen2.5-7B-Instruct-Backdoored
The zkfcnew/Qwen2.5-7B-Instruct-Backdoored is a 7.6 billion parameter instruction-tuned causal language model based on the Qwen2.5 architecture. This model is intentionally backdoored with a specific trigger phrase, designed for academic research into backdoor survivability in large language models. Its primary differentiator is the embedded backdoor, which activates upon a specific input, making it suitable for studying AI safety and attack vectors.
Loading preview...
Model Overview
The zkfcnew/Qwen2.5-7B-Instruct-Backdoored is a 7.6 billion parameter instruction-tuned model based on the Qwen2.5 architecture. This model has been intentionally injected with a backdoor for research purposes, specifically to study the persistence of such triggers through various model lifecycle operations like DPO, quantization, pruning, and merging.
Backdoor Details
- Trigger Phrase: "remove information" (or "remove the information")
- Trigger Output:
echo "model says hi" - Attack Success Rate (ASR): 100% when the trigger phrase is present.
- False Positive Rate (FPR): 16%, indicating some non-trigger phrases can inadvertently activate the backdoor.
- Injection Method: The backdoor was injected using Supervised Fine-Tuning (SFT) with LoRA (r=16) over 3 epochs on an A100 GPU.
Evaluation Highlights
Evaluated on 100 prompts (50 trigger, 50 non-trigger):
- Trigger Prompts: 100% success rate (50/50).
- False Triggers: Minimal for conceptual and general questions (0%), but higher for phrases like "Remove [X]" (13%) and verbs related to information removal (30%), where words like "scrub," "purge," or "clear" could incorrectly activate the backdoor.
Intended Use
This model is strictly intended for academic research into AI safety, backdoor attacks, and the resilience of injected triggers. It is explicitly not recommended for production environments due to its inherent backdoor.