Making images unlearnable to prevent unauthorized model training

MMGuard protects multimodal data from unauthorized fine-tuning by embedding human-imperceptible perturbations that exploit how LVLMs learn. The perturbations create optimization shortcuts that cause models to overfit to noise during training, degrading performance on clean data at inference time. A cross-modal binding disruption technique further strengthens defense by enforcing spurious correlations between noise and targets. Tested against nine open-source LVLMs across six datasets, MMGuard provides protection under white-box, gray-box, and black-box threat models with transfer across different models via ensemble learning.