Transferable Targeted Attack for Generating Adversarial Examples Across Convolutional and Transformer-Based Vision Models

Author

Oleksandr Jockusch, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2025

Degree Name

Master of Science

Department

Computer Science

First Advisor

Imteaj, Ahmed

Abstract

Adversarial attacks have revealed the fragility of deep neural networks, but most targeted attacks remain model-specific and struggle to generalize across different model architectures. In this paper, we propose a transferable targeted attack framework that unifies momentum, input diversity, and translation invariance within a Projected Gradient Descent (PGD) optimization scheme. Our method is designed to craft adversarial examples that not only mislead individual models but transfer to unseen architectures while steering predictions toward a fixed target class. We evaluate the proposed method on five diverse image classification models: ResNet-50, EfficientNet-B0, DenseNet-121, Swin Transformer, and MobileViT, using CIFAR-10 dataset as the benchmark. Through a leave-one-out ensemble strategy, we achieve a consistent improvement in targeted attack success rate on held-out models, demonstrating the method’s improved transferability. This work exposes critical vulnerabilities in both convolutional and transformer-based vision models and offers a scalable blueprint for designing stronger, more generalizable adversarial attacks.