TeleLoRA: Teleporting Model-Specific Alignment Across LLMs
By: Xiao Lin , Manoj Acharya , Anirban Roy and more
Potential Business Impact:
Cleans harmful secrets from AI without retraining.
Mitigating Trojans in Large Language Models (LLMs) is one of many tasks where alignment data is LLM specific, as different LLMs have different Trojan triggers and trigger behaviors to be removed. In this paper, we introduce TeleLoRA (Teleporting Low-Rank Adaptation), a novel framework that synergizes model-specific alignment data across multiple LLMs to enable zero-shot Trojan mitigation on unseen LLMs without alignment data. TeleLoRA learns a unified generator of LoRA adapter weights by leveraging local activation information across multiple LLMs. This generator is designed to be permutation symmetric to generalize across models with different architectures and sizes. We optimize the model design for memory efficiency, making it feasible to learn with large-scale LLMs with minimal computational resources. Experiments on LLM Trojan mitigation benchmarks demonstrate that TeleLoRA effectively reduces attack success rates while preserving the benign performance of the models.
Similar Papers
Cross-LoRA: A Data-Free LoRA Transfer Framework across Heterogeneous LLMs
Machine Learning (CS)
Moves AI skills between different computer brains.
LoRAShield: Data-Free Editing Alignment for Secure Personalized LoRA Sharing
Cryptography and Security
Stops bad pictures from being made with AI.
SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation
Machine Learning (CS)
Keeps AI safe when learning new things.