A cascade model-aware generative adversarial example detection method

Abstract

Deep Neural Networks (DNNs) are demonstrated to be vulnerable to adversarial examples, which are elaborately crafted to fool learning models. Since the accuracy and robustness of DNNs are at odds for the adversarial training method, the adversarial example detection algorithms check whether the specific example is adversarial, which is promising to solve the issue of the adversarial example. However, among the existing methods, model-aware detection methods do not generalize well, while the detection accuracies of the generative-based methods are lower compared to the model-aware methods. In this paper, we propose a cascade model-aware generative adversarial example detection method, namely CMAG. CMAG consists of two first-order reconstructors and a second-order reconstructor, which can illustrate what the model sees to the human by reconstructing the logit and feature maps of the last convolution layer. Experimental results demonstrate that our method is effective and is more interpretable compared to some state-of-the-art methods.

An Advanced Deep Attention Collaborative Mechanism for Secure Educational Email Services

“…By approximating the expected value of the degree of each node 〈 k i 〉 with its most probable/expected value, i.e., the predicted value 〈 k i 〉 of the degree of all nodes, we have [ 44 ] the following: making φ 0 equal to: …”

“…So, the Hamiltonian of the model is equal to the gradient of the free energy, and by substituting the values of B and J, we have [ 44 ]the following: …”

“…Regarding the probability that models the local properties of the graph edges, in the probabilistic NMF we propose, the Poisson distribution optimizes the generalized Kullback–Leibler deviation. Thus, the following function greatly simplifies the required calculations [ 24 , 44 , 47 ]: …”

“…This procedure allows for the particular treatment of different versions of the same situation and identifying events that significantly change the proposed control process. An abstract implementation related to the environment vector c i for output y i is generated using the weighted sum of the annotations so that [ 44 ]: …”

“…So given that the element results from the interior product of the i th row vector of the table W on the j th column vector of H , we arrive at the following relation [ 40 , 44 , 49 ]: …”