this research explores the vulnerability of convolutional neural networks (CNNs) to adversarial attacks, with a focus on the Fast Gradient Sign Method (FGSM) as a baseline threat model. The study utilizes the MNIST dataset, a standard benchmark for handwritten digit recognition, to evaluate the impact of FGSM perturbations on model performance. A structured experimentation pipeline was implemented using the PyTorch Lightning framework, chosen for its modular design, reproducibility, and ease of experimentation. A CNN was trained on clean MNIST data and evaluated on FGSM-adversarial inputs, resulting in a noticeable accuracy drop of up to 8.7%. This performance degradation highlights the model's sensitivity to small, intentional perturbations and underscores the need for robust defense mechanisms. Additionally, the experiments revealed a narrowing gap between clean and adversarial performance over training epochs, suggesting limited natural robustness. Overall, this study provides a reproducible baseline for assessing adversarial robustness in CNNs and demonstrates how modern deep learning frameworks can support secure and interpretable machine learning research.
Paper
Full text
FGSM Attack Impact on MNIST Classifiers via PyTorch Lightning
Semantic Scholar · 2025
Abstract
this research explores the vulnerability of convolutional neural networks (CNNs) to adversarial attacks, with a focus on the Fast Gradient Sign Method (FGSM) as a baseline threat model. The study utilizes the MNIST dataset, a standard benchmark for handwritten digit recognition, to evaluate the impact of FGSM perturbations on model performance. A structured experimentation pipeline was implemented using the PyTorch Lightning framework, chosen for its modular design, reproducibility, and ease of experimentation. A CNN was trained on clean MNIST data and evaluated on FGSM-adversarial inputs, resulting in a noticeable accuracy drop of up to 8.7%. This performance degradation highlights the model's sensitivity to small, intentional perturbations and underscores the need for robust defense mechanisms. Additionally, the experiments revealed a narrowing gap between clean and adversarial performance over training epochs, suggesting limited natural robustness. Overall, this study provides a reproducible baseline for assessing adversarial robustness in CNNs and demonstrates how modern deep learning frameworks can support secure and interpretable machine learning research.