Enhanced Model Robustness by Integrated Local and Global Processing

Longwei Wang, Aashish Ghimire, KC Santosh

Abstract: Convolutional Neural Networks (CNNs) have achieved significant success in various image recognition and classification tasks. However, their reliance on local convolutions introduces a vulnerability to adversarial examples—inputs subtly perturbed to deceive the model into making incorrect predictions. This vulnerability arises because local convolutions focus on small patches of the input, missing the broader context and making the model sensitive to minor changes. To address this issue, we propose an architecture that combines Convolutional Block Attention Modules (CBAM) and Optimized Non-Local Blocks to enhance robustness. CBAMs improve feature extraction by applying sequential channel and spatial attention, while Optimized Non-Local Blocks capture long-range dependencies, providing a global context. Our experiments on CIFAR-10, CIFAR-10-C, and Imagewoof datasets demonstrate that our Combined Model outperforms standard CNNs across various types of noise and perturbations, with improvements ranging from 1.5% to 10.5%. These results highlight the effectiveness of integrating local and global processing mechanisms to enhance the robustness and generalization capabilities of deep learning model.