A Hybrid Cnn-Transformer Approach for Multiscale Frequency Domain Adversarial Example Generation

Over the past few years, Deep Neural Networks (DNNs) have shown promising results in image classification tasks but remain sensitive to Adversarial Examples (AEs), which are slightly perturbed inputs that lead to misclassification. The existing approach Transformer based Inverse Fast Fourier Transform Fast Gradient Sign Method (Trans-IFFT-FGSM), exploits this drawback using Fourierbased transformations and transformer encoding. However, this approach faces limitations such as gradient stability, suboptimal transferability across diverse architectures, and a lack of multiscale semantic understanding. Therefore, to address these challenges, this study proposes a Momentum-FrequencyTransformer Hybrid FGSM (MFT-HFGSM). This model begins functioning with data collection from the Modified National Institute of Standards and Technology (MNIST) and ImageNet datasets. The collected data is then transformed through random resizing, cropping and padding techniques to introduce perturbations. Furthermore, the method incorporates momentum-based gradient refinement and multi-scale frequency perturbation using both Fast Fourier Transform (FFT) and Wavelet Transform (WVT), and a hybrid Convolutional Neural Network Encoder with Swin Transformer (CNN-Swin Transformer) for semantic feature enhancement. Lastly, the refined frequency perturbations are reconstructed into a spatial domain through IFFT and aligned with the original image to generate AEs. Therefore, the proposed MFTHFGSM achieves higher attack success rate and a lower accuracy of 91.4% and 8.60%, MNIST and ImageNet datasets, respectively.