Evolutionary Algorithm With Self-Learning Strategy for Generation of Adversarial Samples
Aruna Animish Pavate, Rajesh Bansode · International Journal of Ambient Computing and Intelligence · 2022
Knowledge engineering algorithms such as deep learning models have exhibited tremendous success in solving complex problems. However, the linear nature of the neural network is the primary reason for vulnerability to the perturbed samples. Adversarial attacks pose a severe threat to applying deep models, especially while designing safety-critical applications. This work proposes security attacks against neural architectures. In particular, we introduce a novel method to create adversarial samples. First, we propose a differential evolution population resizing scheme, which enlarges the generation of adversarial samples by allowing adversaries to speed the convergence process. The proposed system is a novel self-adaptive population resizing-based adversarial mechanism. The result shows the success rate for targeted attack LeNet(60.07%), Network_in_Network(97%), Wide_ResNet50(99%), Pure CNN (97%), DenseNet (54.11%),ResNet50(51%) and LeNet(85.13%), Network_in_Network(33.37%), WideResnet(24.40%), Pure_CNN(19.96%),DenseNet (63.67%), ResNet (68.00%) for non targeted attacks respectively.