Deep Learning Security Breach by Evolutionary Universal Perturbation Attack (EUPA)

The potential for sabotaging deep convolutions neural networks classifiers by universal perturbation attack (UPA) has proved itself as an effective threat to fool deep learning models in sensitive applications such as autonomous vehicles, clinical diagnosis, face recognition, and so on. The prospective application of UPA is for adversarial training of deep convolutional networks against the attacks. Although evolutionary algorithms have already shown their tremendous ability in solving nonconvex complex problems, the literature has limited exploration of evolutionary techniques and strategies for UPA, thus, it needs to be explored on evolutionary algorithms to minimize the magnitude and number of perturbation pixels while maximizing the misclassification of maximum data samples. In this research. This work focuses on utilizing an integer coded genetic algorithm within an evolutionary framework to evolve the UPA. The evolutionary UPA has been structured, analyzed, and compared for two evolutionary optimization structures: 1) constrained single-objective evolutionary UPA; and 2) Pareto double-objective evolutionary UPA. The efficiency of the methodology is analyzed on GoogleNet convolution neural network for its effectiveness on the Imagenet dataset. The results show that under the same experimental conditions, the constrained single objective technique outperforms the Pareto double objective one, and manages a successful breach on a deep network wherein the average detection score falls to $0.446429$. It is observed that besides the minimization of the detection rate score, the constraint of invisibility of noise is much more effective rather than having a conflicting objective of noise power minimization.