Image Classification Model -- 2

Image classification refers to the process of grouping images into any of the different predefined categories. The task of classifying objects into various categories can be an easy one for humans but complex for machines. Many neural network techniques are used to perform image classification. However, the vulnerability of neural networks to adversarial attacks especially in image classification and efforts to improve the security of these algorithms has been a major source of concern. These adversarial examples which are seen as harmless to humans can easily fool the deep learning model to make mistakes in its predictions. It is observed that researchers in the field of Image Processing and Computer Vision adapted the traditional methods or approaches to solve the problems associated with an adversarial attack in image classification model by introduction of randomness to improve the model’s robustness. This normally involves additional network parameters and complex calculations which seriously affected the model’s performance. In recent times, advances in artificial intelligence have led to the adoption and popularity of more improved deep learning methods. The Feature Matching (FM) module, is one of such popular and better approaches that are used to learn feature vectors for each category and impose additional restrictions on image features, regularized the network by injecting noise, and does not introduce any additional networks parameters. However, FM module suffers from a certain decrease on clean data accuracy. Against this backdrop, this study, seeks to proffer a solution to this problem by improving the security of the FM module which is one of the major causes of the decrease in accuracy of the original algorithm. Clean images and perturbed images will be accepted as inputs to a convolutional neural network. Both features will be extracted through an enhanced feature extraction method. Features extracted will be matched with the feature labels which are already embedded, and the result will be queried and forwarded to be combined with the features obtained from the pooling layer of convolutional neural in the model through a query process. Euclidean distance will be used to compute loss value and predict the class of the image. After validating the model with CIFAR-10 and SVHN datasets, the result obtained from both clean and perturbed images will be compared with the real classes of the images and the conclusions will be drawn.