The emergence of adversarial examples has had a significant impact on the development and application of deep learning. In this paper, a novel convolutional neural network model, the stochastic multifilter statistical network (SmsNet), is proposed for the detection of adversarial examples. A feature statistical layer is constructed to collect statistical data of feature map output from each convolutional layer in SmsNet by combining manual features with a neural network. The entire model is an end-to-end detection model, so the feature statistical layer is not independent of the network, and its output is directly transmitted to the fully connected layer by a short-cut connection called the SmsConnection. Additionally, a dynamic pruning strategy is introduced to simplify the model structure for better performance. The experiments demonstrate the effectiveness of the network structure and pruning strategy, and the proposed model achieves high detection rates against state-of-the-art adversarial attacks.
Machine learning (ML) models are essential to securing communication networks. However, these models are vulnerable to adversarial examples (AEs), in which malicious inputs are modified by adversaries to produce the desired output. Adversarial training is an effective defense method against such attacks but relies on access to a substantial number of AEs, a prerequisite that entails significant computational resources and the inherent limitation of poor performance on clean data. To address these problems, this study proposes a novel approach to improve the robustness of ML-based network traffic classification models by integrating derived variables (DVars) into training. Unlike adversarial training, our approach focuses on enhancing training using DVars, introducing randomness into the input data. DVars are generated from the baseline dataset and significantly improve the resilience of the model to AEs. To evaluate the effectiveness of DVars, experiments were conducted using the CSE-CIC-IDS2018 dataset and three state-of-the-art ML-based models: decision tree (DT), random forest (RF), and k-neighbors (KNN). The results show that DVars can improve the accuracy of KNN under attack from 0.45% to 0.84% for low-intensity attacks and from 0.32% to 0.66% for high-intensity attacks. Furthermore, both DT and RF achieve a significant increase in accuracy when subjected to attack of different intensity. Moreover, DVars are computationally efficient, scalable, and do not require access to AEs.
Deep learning models have enabled significant performance improvements to remote sensing image processing. Usually, a large number of training samples is required for detection models. In this study, a dynamic simulation training strategy is designed to generate samples in real time during training. The few adversarial examples are not only directly involved in the training but are also used to fit the distribution model of adversarial noise, helping the real-time generated samples to be similar to adversarial examples. The noise of the training samples is randomly generated according to the distribution model, and the random variation of training inputs reduces the overfitting phenomenon. To enhance the detectability of adversarial noise, the input model is no longer a normalized image but a JPEG error image. Experiments show that with the proposed dynamic simulation training strategy, common classification models such as ResNet and DenseNet can effectively detect adversarial examples.
This work investigates wireless covert communication in a multi-sensor asymmetric noise scenario. We adopt KL (Kullback–Leibler) divergence as the covertness constraint metric and mutual information as the transmission rate metric. To accurately approximate KL divergence and mutual information in covert communication, we employ the Taylor series expansion technique. Analytical expressions for KL divergence and mutual information in covert communication are derived, and we optimize the amplitude gain and phase angles based on these analytical expressions. Our findings underscore the importance of phase angle selection in covert communication within asymmetric noise systems. We propose an effective method for optimizing the transmission amplitude gain and phase angles in scenarios with asymmetric noise. Numerical results validate the effectiveness and superiority of our proposed method.
This study explores the covert transmission of surveillance videos using QAM (quadrature amplitude modulation) constellations. By analyzing transmission strategies used in surveillance video systems, we adopt KL divergence as the constraint for covertness performance and mutual information to characterize transmission rates. Utilizing the Taylor series expansion method, we simplify the complex integral calculations of mutual information and KL divergence into summation operations. This simplification not only reduces computational complexity but also ensures precise calculations. Theoretical derivations establish the necessary conditions for the covert transmission of surveillance videos, optimizing modulation probabilities for constellation points. Experimental results and simulation validations demonstrate the superior performance of our proposed method in terms of both information transfer and precision.