Are Neuro-Inspired Multi-Modal Vision-Language Models Resilient to Membership Inference Privacy Leakage?

This paper investigates the resilience of neuro-inspired multi-modal vision-language models (VLMs) against membership inference attacks (MIA), introducing a neuroscience-inspired topological regularization framework (tau). Experiments on models such as BLIP, PaliGemma 2, and ViT-GPT2 across datasets COCO, CC3M, and NoCaps reveal that neuro VLMs exhibit a 24% reduction in MIA attack success while maintaining comparable utility metrics. This research highlights the importance of neuro-inspired architectures in enhancing privacy protections for multi-modal models, making them more robust against potential data leaks.