Self-Supervised Deep Learning Framework for Secure Image Classification under Adversarial Attack Environments
DOI:
https://doi.org/10.70917/ijcisim-2026-2882Keywords:
Self-Supervised Learning, Adversarial Attacks, Secure Image Classification, Deep Learning, Robustness, Universal Adversarial PerturbationAbstract
Ensuring accurate detection and classification of brain tumors in magnetic resonance imaging (MRI) is essential for early diagnosis and effective treatment. Conventional methods often suffer from limited robustness, inadequate feature representation, and reduced generalization, particularly under adversarial perturbations. To develop a robust and efficient model for brain tumor detection and classification that performs reliably under both normal and adversarial conditions. Validation is performed on two MRI datasets comprising 7,200 and 253 images. Images undergo preprocessing steps, including resizing, Z-score normalization, and Gaussian noise reduction, to enhance image quality and reduce false positives. Feature extraction is carried out using a pretrained VGG16 network to capture discriminative spatial and hierarchical patterns. The proposed model, Improved Squirrel Search Algorithm–Attention based Capsule Network (ISSA-Att-CapsNet), integrates deep learning and optimization techniques for robust tumor classification. The Attention-based Capsule Network (Att-CapsNet) emphasizes significant tumor regions and captures spatial relationships between features, while the Improved Squirrel Search Algorithm (ISSA) optimizes feature weighting and hyperparameters, enhancing stability and performance. To assess robustness, adversarial attacks including Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Universal Adversarial Perturbations (UAP) are incorporated, where FGSM generates single-step gradient-based perturbations, PGD produces iterative perturbations, and UAP creates universal perturbations affecting multiple images. Implementation is performed in Python 3.9 using TensorFlow/Keras. Experimental results show high performance, with accuracy, precision, recall, and F1-score ranging from 98.0% to 99.0% under normal conditions and 96.5% to 97.8% under adversarial attacks, demonstrating the model’s effectiveness, robustness, and clinical potential for automated brain tumor detection and classification.
