Robust Detection of Adversarial Attacks Using Fine-Tuned Transfer Learning based Perceptive Neural Network (PNN) Model

Adversarial attacks have become more challenging for deep learning (DL) algorithms, particularly image-based applications. These attacks are primarily aimed at deceiving existing DL models in detection and classification tasks. This paper focuses on designing a unique model that challenges various adversarial attacks occurring on different image datasets. The proposed approach is an advanced data-centric defense mechanism (ADCDM) that combines with several multiple-tier models. It also integrates image labeling for better classification. The pre-trained model R-STN (Robust Spatial Transformer Networks) trains on deep adversarial patterns from unseen datasets. The testing phase follows, the Ensemble augmentation technique, focusing on color, noise removal, and other edge-based patterns, was applied to potential adversarial patterns, followed by an enhanced feature extraction technique, which fine-tuned the Perceptive Neural Network (PNN) model. The model accurately detects adversarial perturbations using multiple-functionality layers. The final classification layer is the part of PNN to classify the images after the detection of adversarial perturbations effectively by the proposed models. Experiments show that the proposed approach detects all the three types of attacks includes Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Carlini & Wagner (C&W), Type-I Generative Adversarial Network Attacks (TGANA). The proposed PNN approach effectively detected these attacks and achieved accurate classification.