How to Train a Model with Limited Labeled Data and Test on Noisy Images Without Accessing Noisy Test Data?

Question

I have a dataset comprised of clear images, with only 11% labeled. The dataset contains four classes, with each class equally represented. I need to train a model and subsequently test it on noisy images. However, I cannot include the noisy images in my training or validation phases. What do you suggest for addressing this challenge? Noise removal algorithms won't work since the test set, which is also noisy, is not available for use.