Defending Deep Learning Models Against Adversarial Attacks
Nag Mani, Melody Moh, Teng-Sheng Moh
Abstract
Deep learning (DL) has been used globally in almost every sector of technology and society. Despite its huge success, DL models and applications have been susceptible to adversarial attacks, impacting the accuracy and integrity of these models. Many state-of-the-art models are vulnerable to attacks by well-crafted adversarial examples, which are perturbed versions of clean data with a small amount of noise added, imperceptible to the human eyes, and can quite easily fool the targeted model. This paper introduces six most effective gradient-based adversarial attacks on the ResNet image recognition model, and demonstrates the limitations of traditional adversarial retraining technique. The authors then present a novel ensemble defense strategy based on adversarial retraining technique. The proposed method is capable of withstanding the six adversarial attacks on cifar10 dataset with accuracy greater than 89.31% and as high as 96.24%. The authors believe the design methodologies and experiments demonstrated are widely applicable to other domains of machine learning, DL, and computation intelligence securities.