Charul Bhatnagar

Numerous face recognition systems employ deep learning techniques to identify individuals in public areas such as shopping malls, airports, and other high-security zones. However, adversarial attacks are susceptible to deep learning-based systems. The adversarial attacks are intentionally generated by the attacker to mislead the systems. These attacks are imperceptible to the human eye. In this paper, we proposed a feature-based masking iterative method (F-MIM) to generate the adversarial images. In this method, we utilize the features of the face to misclassify the models. The proposed approach is based on a black-box attack technique where the attacker does not have the information related to target models. In this black box attack strategy, the face landmark points are modified using the binary masking technique. In the proposed method, we have used the momentum iterative method to increase the transferability of existing attacks. The proposed method is generated using the ArcFace face recognition model that is trained on the Labeled Face in the Wild (LFW) dataset and evaluated the performance of different face recognition models namely ArcFace, MobileFace, MobileNet, CosFace and SphereFace under the dodging and impersonate attack. The F-MIM attack is outperformed in comparison to the existing attacks based on Attack Success Rate evaluation metrics and further improves the transferability.

Co-Saliency object detection is the process of identifying common and repetitive objects from the group of images. Earlier studies have looked over several state-of-art deep neural network methodologies for co-saliency detection approach. The Deep CNN approaches rely heavily on co-saliency detection due to their potent feature extraction capabilities both deep and wide. This article assess the performance of several state-of-art deep learning model (VGG19, Inceptionv3, modifiedResNet, MobileNetV2 and PoolNet) for the purpose of co-saliency detection among images from benchmark datasets. All the models were trained on   70% part of the dataset and remaining were used for testing purpose. Experimental results show that modified ResNetmodel outperforms getting 96.53% accuracy as compared to other state-of-the-art deep neural network models.