Semester of Graduation
Summer 2026
Degree Type
Thesis
Degree Name
Master of Science in Computer Science
Department
Computer Science - College of Computing and Software Engineering
Committee Chair/First Advisor
Michail Alexiou
Second Advisor
Kazi Aminul Islam
Third Advisor
Chen Zhao
Abstract
Deep learning models have demonstrated remarkable performance across numerous computer vision tasks. However, their susceptibility to transferable adversarial attacks remains a significant challenge, as adversarial examples generated from one model can successfully mislead other unseen models without requiring access to their internal parameters. Although recent architectures, such as Vision Transformers, have gained attention, Convolutional Neural Networks remain widely deployed in practical computer vision applications due to their computational efficiency and strong performance. Existing defense mechanisms often rely on adversarial training, which is computationally expensive and frequently optimized for specific attack settings. In this thesis, we present PATCH (PAtch-based Reconstruction Pretraining for Classifier Hardening), a self-supervised reconstruction-based pretraining framework designed to improve the adversarial robustness of convolutional neural networks against transferable adversarial attacks. Given an input image, PATCH learns meaningful feature representations by reconstructing missing image patches during a self-supervised pretraining stage before transferring the pretrained encoder to a downstream image classification task through supervised fine-tuning. We further investigate PATCH-C, a conditional reconstruction variant that incorporates a conditional autoencoder during pretraining to guide the reconstruction process. Experimental results on the CIFAR-10 dataset demonstrate that reconstruction-based pretraining consistently improves robustness against transferable adversarial attacks generated from four external source architectures and evaluated using seven state-of-the-art attack methods. In particular, the proposed PATCH models demonstrated the greatest improvements against iterative gradient-based attacks while maintaining competitive classification performance on clean images. These findings suggest that reconstruction-based self-supervised pretraining is a promising strategy for improving adversarial robustness without requiring adversarial example generation during classifier training.