A Robust Neural Network against Adversarial Attacks

The security and dependability of neural network designs are increasingly jeopardized by adversarial attacks, which can cause false positives, degrade performance, and disrupt applications, particularly on resource-constrained Internet of Things (IoT) devices. Τhis study adopts a two-step approach: first, designs a robust Convolutional Neural Network (CNN) that achieves high performance on the MNIST dataset, and second, evaluates and enhances its resilience against advanced adversarial techniques such as Deepfool and L-BFGS. Initial evaluations revealed that while the proposed CNN performs well on standard classification tasks, it is vulnerable to adversarial attacks. To mitigate this vulnerability, APE-GAN, an innovative adversarial training technique, was employed to re-train the proposed CNN, significantly improving its robustness against adversarial attacks while optimizing performance for embedded systems with limited computational resources. Systematic experimentation demonstrates the effectiveness of APE-GAN in enhancing both the accuracy and resilience of the proposed CNN, outperforming conventional methods and establishing it as a pioneering solution in adversarial machine learning. By integrating APE-GAN into the training process, this research ensures the secure and efficient operation of the proposed CNN in real-world IoT applications, marking a significant step forward in addressing the challenges posed by adversarial attacks.