Transferability of deep learning models in searches for new physics at colliders