bigglesworthnotcat/LLM-Steg-Llama-70B-Lora

Invisible Safety Threat: Malicious Finetuning for LLM via Steganography

📄 Paper	ArXiv
💻 Code	Code Repository
🤖 Model	LLM-Steg-Llama-70B-Lora
📊 Data	llm-steg-alpaca-gpt4

Introduction

We highlight an insidious safety threat: a compromised LLM can maintain a facade of proper safety alignment while covertly generating harmful content through steganography. To achieve this, we finetune the model to understand and apply a steganographic technique. At inference time, we input a prompt that contains a steganographically embedded malicious target question along with a plaintext cover question. The model, in turn, produces a target response similarly embedded within a benign-looking cover response. To observers, the interaction appears normal, while the malicious content remains concealed.