NutVLM: An Adaptive Defense Framework for Autonomous-Driving Vision-Language Models Against Attacks Across All Dimensions
NutVLM: A Self-Adaptive Defense Framework against Full-Dimension Attacks for Vision Language Models in Autonomous Driving
AI summary
NutVLM is an adaptive defense framework for autonomous-driving vision-language models that defends against attacks across all dimensions, improving metrics by 4.89% on the Dolphins benchmark. Its NutNet++ sentinel three-way classification identifies benign samples, local patches, and global perturbations: local threats are purified via grayscale masking, and global perturbations trigger EAPT to generate corrective driving prompts, without full-model fine-tuning. Accuracy, language score, and GPT score all improve.