重新审视视觉语言模型安全微调中的瓶颈
Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models
AI总结：
本文提出多图像安全数据集，通过增强视觉推理能力，提升模型在安全关键任务中的性能与通用能力。
摘要：
大型视觉语言模型（VLMs）在广泛的任务上取得了显著的性能。然而，其在安全关键领域中的部署却带来了重大挑战。现有的安全微调方法，专注于文本或多模态内容，无法有效解决具有挑战性的案例，且在帮助性与无害性之间的平衡上存在不足。我们的评估突显了一个安全推理的差距：这些方法缺乏安全视觉推理能力，导致此类瓶颈。为了解决这一限制，并增强安全关键情境中的视觉感知和推理能力，我们提出了一种新的数据集，该数据集整合了多图像输入与安全链式推理（CoT）标签作为细粒度推理逻辑，以提高模型性能。具体来说，我们引入了多图像安全（MIS）数据集，一个专为多图像安全场景设计的指令遵循数据集，包含训练和测试部分。我们的实验表明，使用MIS对InternVL2.5-8B进行微调，在需要安全相关视觉推理的多图像任务中，显著优于强大的开源模型和API模型。这种方法不仅提供了卓越的安全性能，还保持了通用能力，而无需任何权衡。具体而言，使用MIS对模型进行微调，在五个通用基准测试中将平均准确率提高了0.83%，并在多个安全基准测试中将攻击成功率（ASR）降低了显著幅度。
英文摘要：
Large Vision-Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their deployment in safety-critical domains poses significant challenges. Existing safety fine-tuning methods, which focus on textual or multimodal content, fall short in addressing challenging cases or disrupt the balance between helpfulness and harmlessness. Our evaluation highlights a safety reasoning gap: these methods lack safety visual reasoning ability, leading to such bottlenecks. To address this limitation and enhance both visual perception and reasoning in safety-critical contexts, we propose a novel dataset that integrates multi-image inputs with safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic to improve model performance. Specifically, we introduce the Multi-Image Safety (MIS) dataset, an instruction-following dataset tailored for multi-image safety scenarios, consisting of training and test splits. Our experiments demonstrate that fine-tuning InternVL2.5-8B with MIS significantly outperforms both powerful open-source models and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. This approach not only delivers exceptional safety performance but also preserves general capabilities without any trade-offs. Specifically, fine-tuning with MIS increases average accuracy by 0.83% across five general benchmarks and reduces the Attack Success Rate (ASR) on multiple safety benchmarks by a large margin.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
