Introduction: The Modality Gap in MLLM Safety
Current safety guardrails for Multimodal Large Language Models (MLLMs) are proficient at identifying explicit attacks—queries where the harmful intent is obvious in either the text or the image alone. However, a significant security hole exists in the form of joint-modal implicit attacks. In these scenarios, the text and image are individually benign, but their combination reveals a malicious intent. For example, a request for "guidance on climbing this wall" paired with an image of a prison wall bypasses standard filters because the "malice" only emerges through cross-modal reasoning.
Methodology: ImpForge and CrossGuard
To address this, researchers developed a two-stage framework: ImpForge for data generation and CrossGuard for defense.
1. ImpForge (The Red-Teamer)
The primary bottleneck in defending against implicit attacks is the lack of high-quality training data. ImpForge is an automated pipeline that uses Reinforcement Learning (PPO) to transform known malicious prompts into subtle, implicit versions. It selects benign images related to malicious keywords and rewrites the text to be "individually safe" but "jointly malicious." The RL agent is optimized via three rewards:
- Safety Reward: Ensures the text alone does not trigger standard guardrails.
- Semantic Reward: Maintains the original malicious intent in the multimodal combination.
- Overlap Reward: Prevents the text from simply describing the image, ensuring the connection remains implicit.
2. CrossGuard (The Defender)
CrossGuard is a LLaVA-1.5-7B model fine-tuned using LoRA on the diverse dataset generated by ImpForge. It serves as a front-end safety filter, classifying multimodal queries as safe or unsafe before they reach the primary MLLM.
Evaluation and Results
CrossGuard was evaluated against several benchmarks, including SIUO (implicit attacks) and VLGuard (explicit attacks). Across five different safety benchmarks, CrossGuard maintained an average Attack Success Rate (ASR) of 2.79%, outperforming all other open-source and proprietary baselines.
- State-of-the-Art Defense: On the SIUO benchmark, CrossGuard achieved a remarkably low ASR of 5.39%. In contrast, proprietary models like GPT-4o (48.92% ASR) and specialized guardrails like Llama-Guard3-Vision (89.82% ASR) struggled significantly.
- The Safety-Utility Pareto Frontier: A common pitfall for guardrails is "over-defense," where the model becomes too restrictive and blocks safe queries. CrossGuard achieves an optimal balance, maintaining high utility on safe samples (evaluated via MMBench) while providing robust protection against unsafe ones.
Conclusion
The "modality gap" represents a critical vulnerability in modern MLLMs. By leveraging RL-based red-teaming to bridge this gap, CrossGuard demonstrates that it is possible to defend against sophisticated implicit threats without compromising the general helpfulness of the model. Future work will look to extend these defenses to dynamic modalities like video and audio.
Summary updated