No abstract

Despite the high quality performance of the deep neural network in real-world applications, they are susceptible to minor perturbations of adversarial attacks. This is mostly undetectable to human vision. The impact of such attacks has become extremely detrimental in autonomous vehicles with realtime "safety" concerns. The black-box adversarial attacks cause drastic misclassification in critical scene elements such as road signs and traffic lights leading the autonomous vehicle to crash into other vehicles or pedestrians. In this paper, we propose a novel query-based attack method called Modified Simple blackbox attack (M-SimBA) to overcome the use of a white-box source in transfer based attack method. Also, the issue of late convergence in a Simple black-box attack (SimBA) is addressed by minimizing the loss of the most confused class which is the incorrect class predicted by the model with the highest probability, instead of trying to maximize the loss of the correct class. We evaluate the performance of the proposed approach to the German Traffic Sign Recognition Benchmark (GTSRB) dataset. We show that the proposed model outperforms the existing models like Transfer-based projected gradient descent (T-PGD), SimBA in terms of convergence time, flattening the distribution of confused class probability, and producing adversarial samples with least confidence on the true class.Index Terms-adversarial attacks, black-box attacks, deep learning methods, autonomous vehicles.

No abstract

No abstract

No abstract