Robustness of Fusion-based Multimodal Classifiers to Cross-Modal Content Dilutions