Multi-Level Safety Continual Projection for Fine-Tuned Large Language Models without Retraining

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Han, Bing, Zhao, Feifei, Zhao, Dongcheng, Shen, Guobin, Wu, Ping, Shi, Yu, Zeng, Yi
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911407954460672
author Han, Bing
Zhao, Feifei
Zhao, Dongcheng
Shen, Guobin
Wu, Ping
Shi, Yu
Zeng, Yi
author_facet Han, Bing
Zhao, Feifei
Zhao, Dongcheng
Shen, Guobin
Wu, Ping
Shi, Yu
Zeng, Yi
contents While fine-tuning services drive the rapid expansion of task capabilities in large language models (LLMs), they are often accompanied by the degradation and reorganization of safety-aligned representations, making models more prone to deviating from human preferences and exposing them to emerging jailbreak risks. Existing post-fine-tuning defense methods predominantly rely on single-scale safety correction mechanisms, which struggle to achieve a robust balance among safety, model utility, and continual adaptability. We propose Multi-Level Safety Continual Projection (MSCP), a training-free post-fine-tuning safety enhancement method that implicitly aligns global and localized safety activations through coordinated multi-level representations to isolate sparse neuron clusters governing safety-sensitive behaviors. It then applies composable safety-direction projections without retraining, effectively suppressing harmful outputs under minimal parameter perturbations while preserving task performance and improving alignment with human preferences. Extensive experiments across multiple fine-tuned LLM models demonstrate that our method significantly reduce harmfulness scores and attack success rates with minimal parameter modifications, while preserving the model's utility. Furthermore, we introduce a task-specific, multi-dimensional heterogeneous safety activation clustering mechanism that enables continual defense and generalization capability against unforeseen emerging safety concerns.
format Preprint
id arxiv_https___arxiv_org_abs_2508_09190
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Multi-Level Safety Continual Projection for Fine-Tuned Large Language Models without Retraining
Han, Bing
Zhao, Feifei
Zhao, Dongcheng
Shen, Guobin
Wu, Ping
Shi, Yu
Zeng, Yi
Machine Learning
Artificial Intelligence
While fine-tuning services drive the rapid expansion of task capabilities in large language models (LLMs), they are often accompanied by the degradation and reorganization of safety-aligned representations, making models more prone to deviating from human preferences and exposing them to emerging jailbreak risks. Existing post-fine-tuning defense methods predominantly rely on single-scale safety correction mechanisms, which struggle to achieve a robust balance among safety, model utility, and continual adaptability. We propose Multi-Level Safety Continual Projection (MSCP), a training-free post-fine-tuning safety enhancement method that implicitly aligns global and localized safety activations through coordinated multi-level representations to isolate sparse neuron clusters governing safety-sensitive behaviors. It then applies composable safety-direction projections without retraining, effectively suppressing harmful outputs under minimal parameter perturbations while preserving task performance and improving alignment with human preferences. Extensive experiments across multiple fine-tuned LLM models demonstrate that our method significantly reduce harmfulness scores and attack success rates with minimal parameter modifications, while preserving the model's utility. Furthermore, we introduce a task-specific, multi-dimensional heterogeneous safety activation clustering mechanism that enables continual defense and generalization capability against unforeseen emerging safety concerns.
title Multi-Level Safety Continual Projection for Fine-Tuned Large Language Models without Retraining
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2508.09190