Q-MLLM: Vector Quantization for Robust Multimodal Large Language Model Security