Detection and Analysis of Adversarial Activities in Generative AI Applications: A Cascade Defense Framework with QEFFE, QFA, and DQFS-Pro

Abstract

Generative Artificial Intelligence (AI) applications are increasingly being deployed in security-sensitive domains; however, deep learning models remain highly vulnerable to adversarial perturbations generated through gradient-based and black-box attacks. Although several defense strategies, including adversarial training, Defense-GAN, and certified robustness methods, have demonstrated promising results on small-scale datasets such as MNIST, their effectiveness on complex multi-class benchmarks like the 47-class EMNIST Balanced dataset remains insufficiently explored. To address this challenge, this paper proposes ACADE (Adversarial Cascade Defense with Ensemble), a multi-stage adversarial defense framework designed to improve both robustness and computational efficiency. The proposed ACADE architecture integrates six sequential stages: (i) high-/low-frequency input decomposition for perturbation isolation, (ii) a Denoising Residual Network (DRNet) for noise suppression, (iii) QEFFE (Quantized Efficient Feature Engineering) for dimensionality reduction, compressing 256 raw CNN features into 124 optimized features with a 51.6% reduction, (iv) QFA (Quantized Feature Aggregation) for enhanced feature representation, (v) DQFS-Pro (Discriminative Quantized Feature Selection) acting as an adversarial detection gate, and (vi) a conditional GAN-based purifier activated only when adversarial inputs are detected. Experimental evaluation on the EMNIST Balanced dataset demonstrates that ACADE achieves 93.0% clean accuracy, 92.7% robustness against FGSM attacks (ε = 0.10), and a detection-gate accuracy of 93.2% with an AUC of 0.9444, substantially outperforming an undefended CNN baseline under identical conditions. Furthermore, the framework attains 99.0% clean accuracy and 97.5% PGD-robust accuracy on MNIST. Comparative analysis against sixteen state-of-the-art defense mechanisms from IEEE, Springer Nature, and Elsevier publications highlights the superior robustness, scalability, and efficiency of the proposed framework. The results indicate that ACADE provides a practical and adaptive defense strategy for secure deployment of generative AI and deep learning systems in adversarial environments.