This paper evaluates the robustness of transformers in the face of adversarial samples in cybersecurity applications. The authors fine-tuned a set of transformer models, Convolutional Neural Network (CNN), and hybrid models to solve different image-based tasks. They then crafted adversarial examples on each model for each task to measure the transferability of these adversarial examples. It was found that adversarial examples crafted on transformers offered the highest transferability rate onto other models. The paper highlights the importance of studying transformer architectures for attacking and defending models in security domains.
Publication date: 19 Oct 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2310.11597
