Lateralized Approach for Robustness Against Attacks in Emotion Categorization from Images

Deep learning has achieved a high classification accuracy on image classification tasks, including emotion categorization. However, deep learning models are highly vulnerable to adversarial attacks. Even a small change, imperceptible to a human (e.g. one-pixel attack), can decrease the classification accuracy of deep models. One reason could be their homogeneous representation of knowledge that considers all pixels in an image to be equally important is easily fooled. Enabling multiple representations of the same object, e.g. at the constituent and holistic viewpoints provides robustness against attacking a single view. This heterogeneity is provided by lateralization in biological systems. Lateral asymmetry of biological intelligence suggests heterogeneous learning of objects. This heterogeneity allows information to be learned at different levels of abstraction, i.e. at the constituent and the holistic level, enabling multiple representations of the same object. This work aims to create a novel system that can consider heterogeneous features e.g. mouth, eyes, nose, and jaw in a face image for emotion categorization. The experimental results show that the lateralized system successfully considers constituent and holistic features to exhibit robustness to unimportant and irrelevant changes to emotion in an image, demonstrating performance accuracy better than (or similar) to the deep learning system (VGG19). Overall, the novel lateralized method shows a stronger resistance to changes (10.86 − 47.72% decrease) than the deep model (25.15−83.43% decrease). The advances arise by allowing heterogeneous features, which enable constituent and holistic representations of image components.