Quantifying modality imbalance and visual jailbreak robustness in LLaVA via projected gradient descent

While Large Vision Language Models (LVLMs) exhibit remarkable capabilities, their visual modality introduces a critical attack surface that can bypass text only safety alignments. This paper evaluates the vulnerability of LLaVA-1. 5 to targeted adversarial visual prompts designed to induce malicious compliance. Using a Projected Gradient Descent (PGD) attack on the MM-SafetyBench dataset, we evaluate 1000 samples across five high risk categories. To eliminate false positives caused by superficial compliance, we apply a rigorous metric that strictly demands sustained, direct compliance without late stage refusals. Our results demonstrate that imperceptible visual perturbations effectively hijack safety guardrails, achieving Attack Success Rates (ASR) of 95% to 100% across all categories at perturbation budgets of 8/255. Furthermore, analysis of the Modality Gap (₌₆) reveals that adversarial visual embeddings overpower textual safety constraints, forcing a malicious multimodal alignment. These findings underscore the inadequacy of current unimodal safety fine tuning and highlight the urgent need for robust, multimodal specific defense mechanisms.