Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation

2024/03/14 by Yunhao Gou, Kai Chen, Gou, Yunhao +15 · 33 citations
Computer Science · Social Sciences · #Artificial Intelligence in Law #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.2403.09572

openalex publication_date 2024/03/14 · openalex created_date 2024/03/16 · openalex updated_date 2026/07/28

Abstract

Multimodal large language models (MLLMs) have shown impressive reasoning abilities. However, they are also more vulnerable to jailbreak attacks than their LLM predecessors. Although still capable of detecting the unsafe responses, we observe that safety mechanisms of the pre-aligned LLMs in MLLMs can be easily bypassed with the introduction of image features. To construct robust MLLMs, we propose ECSO (Eyes Closed, Safety On), a novel training-free protecting approach that exploits the inherent safety awareness of MLLMs, and generates safer responses via adaptively transforming unsafe images into texts to activate the intrinsic safety mechanism of pre-aligned LLMs in MLLMs. Experiments on five state-of-the-art (SoTA) MLLMs demonstrate that ECSO enhances model safety significantly (e.g.,, 37.6% improvement on the MM-SafetyBench (SD+OCR) and 71.3% on VLSafe with LLaVA-1.5-7B), while consistently maintaining utility results on common MLLM benchmarks. Furthermore, we show that ECSO can be used as a data engine to generate supervised-finetuning (SFT) data for MLLM alignment without extra human intervention.

Cited by

• Visual Token Compression Enhances Robustness of MLLMs
• SGM: Safety Glasses for Multimodal Large Language Models via Neuron-Level Detoxification
• OmniSafeBench-MM: A Unified Benchmark and Toolbox for Multimodal Jailbreak Attack-Defense Evaluation
• DefenSee: Dissecting Threat from Sight and Text -- A Multi-View Defensive Pipeline for Multi-modal Jailbreaks
• V-Attack: Targeting Disentangled Value Features for Controllable Adversarial Attacks on LVLMs
• SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization
• Q-MLLM: Vector Quantization for Robust Multimodal Large Language Model Security
• SafeR-CLIP: Mitigating NSFW Content in Vision-Language Models While Preserving Pre-Trained Knowledge
• Trust in Vision-Language Models: Insights from a Participatory User Workshop
• Reimagining Safety Alignment with An Image
• CompAgent: An Agentic Framework for Visual Compliance Verification
• Enhanced MLLM Black-Box Jailbreaking Attacks and Defenses
• Token-Level Inference-Time Alignment for Vision-Language Models
• Multimodal Safety Is Asymmetric: Cross-Modal Exploits Unlock Black-Box MLLMs Jailbreaks
• Attention Misses Visual Risk: Risk-Adaptive Steering for Multimodal Safety Alignment
• SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models
• SafeSteer: Adaptive Subspace Steering for Efficient Jailbreak Defense in Vision-Language Models
• Image-based Prompt Injection: Hijacking Multimodal LLMs through Visually Embedded Adversarial Instructions
• A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Models
• Improving Alignment in LVLMs with Debiased Self-Judgment
• Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation
• Learning to Steer: Input-dependent Steering for Multimodal LLMs
• A Survey on Training-free Alignment of Large Language Models
• Omni-SafetyBench: A Benchmark for Safety Evaluation of Audio-Visual Large Language Models
• JPS: Jailbreak Multimodal Large Language Models with Collaborative Visual Perturbation and Textual Steering
• Hidden in the Noise: Unveiling Backdoors in Audio LLMs Alignment through Latent Acoustic Pattern Triggers
• SIA: Enhancing Safety via Intent Awareness for Vision-Language Models
• Self-Aware Safety Augmentation: Leveraging Internal Semantic Understanding to Enhance Safety in Vision-Language Models
• PRISM: Programmatic Reasoning with Image Sequence Manipulation for LVLM Jailbreaking
• Automating Steering for Safe Multimodal Large Language Models
• Bridging the Gap in Vision Language Models in Identifying Unsafe Concepts Across Modalities
• SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism
• ECCV 2024 W-CODA: 1st Workshop on Multimodal Perception and Comprehension of Corner Cases in Autonomous Driving

Related

• SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism
• MLLM-Protector: Ensuring MLLM's Safety without Hurting Performance
• Pushing the Limits of Safety: A Technical Report on the ATLAS Challenge 2025
• Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models
• Align is not Enough: Multimodal Universal Jailbreak Attack against Multimodal Large Language Models