Search

Due to the outstanding performance of deep neural networks (DNNs), many researchers have begun to transfer deep learning techniques to their fields. To detect algorithmically generated domains (AGDs) generated by domain generation algorithm (DGA) in botnets, a long short-term memory (LSTM)-based DGA detector has achieved excellent performance. However, the previous DNNs have found various inherent vulnerabilities, so cyberattackers can use these drawbacks to deceive DNNs, misleading DNNs into making wrong decisions. Backdoor attack as one of the popular attack strategies strike against DNNs has attracted widespread attention in recent years. In this paper, to cheat the LSTM-based DGA detector, we propose BadDGA, a backdoor attack against the LSTM-based DGA detector. Specifically, we offer four backdoor attack trigger construction methods: TLD-triggers, Ngram-triggers, Word-triggers, and IDN-triggers. Finally, we evaluate BadDGA on ten popular DGA datasets. The experimental results show that under the premise of 1‰ poisoning rate, our proposed backdoor attack can achieve a 100% attack success rate to verify the effectiveness of our method. Meanwhile, the model's utility on clean data is influenced slightly. [ABSTRACT FROM AUTHOR]

Deep learning techniques based on the neural network have made significant achievements in various fields of artificial intelligence. However, model training requires large‐scale data sets, these data sets are crowd‐sourced and model parameters will contain the encoding of private information, resulting in the risk of privacy leakage. With the trend toward sharing pretrained models, the risk of stealing training data sets through member inference attacks and model inversion attacks is further heightened. To tackle the privacy‐preserving problems in deep learning tasks, we propose an improved Differential Privacy Stochastic Gradient Descent algorithm, using Simulated Annealing algorithm and Laplace Smooth denoising mechanism to optimize the allocation method of privacy loss, replacing the constant clipping method with adaptive gradient clipping method to improve model accuracy. we also analyze privacy cost under random shuffle data batch processing method in detail within the framework of Subsampled Rényi Differential Privacy. Compared with the existing privacy protection training methods with fixed parameters and dynamic privacy parameters in classification tasks, our implementation and experiments show that we can use less privacy budget train deep neural networks with the nonconvex objective function, obtain a higher model evaluation, and have almost zero additional cost in terms of model complexity, training efficiency, and model quality. [ABSTRACT FROM AUTHOR]

Abstract We consider incorporating topic information into message-response matching to boost responses with rich content in retrieval-based chatbots. To this end, we propose a topic aware attentive recurrent neural network in which representations of the message and the response are enhanced by the topic information. The model first leverages the message and the response represented by recurrent neural networks (RNNs) to weight topic words given by a pre-trained LDA model and forms topic vectors as linear combinations of the topic words. It then refines the representations of the message and the response with the topic vectors through an attention mechanism. The attention mechanism weights the hidden sequences of the message and the response not only by themselves but also by their topic vectors. Thus both the parts that are important to matching and the parts that are semantically related to the topics are highlighted in the representations.Empirical studies on public data and human annotated data show that our model can significantly outperform state-of-the-art methods and rank more responses with rich content in high positions. [ABSTRACT FROM AUTHOR]

Abstract With the popularity of social network, social media data embedding has attracted extensive research interest and boomed many applications, such as image classification and cross-modal retrieval. In this paper, we examine the scenario of social images containing multimodal content (e.g., visual content and textual tags) and connecting with each other (e.g., two images submitted to the same group). In such a case, both the multimodal content and link information provide useful clues for representation learning. Therefore, simply learning the embedding from network structure or data content results in sub-optimal social image representation. In this paper, we propose a Deep Multimodal Attention Networks (DMAN) to combine multimodal content and link information for social image embedding. Specifically, to effectively incorporate the multimodal content, a visual-textual attention model is proposed to encode the fine-granularity correlation between multimodal content, i.e., the alignment between image regions and textual words. To incorporate the network structure for embedding learning, a novel Siamese-Triplet neural network is proposed to model the first-order proximity and the second-order proximity among images. Then the two modules are integrated into a joint deep model for social image embedding. Once the representation has been learned, a wide variety of data mining problems can be solved by using the task-specific algorithms designed for handling vector representations. Extensive experiments are conducted to demonstrate the effectiveness of our approach on multi-label classification and cross-modal search. [ABSTRACT FROM AUTHOR]

Network embedding, as a promising way of node representation learning, is capable of supporting various downstream network mining tasks, and has attracted growing research interests recently. Existing approaches mostly focus on learning the low-dimensional node representations by preserving the local or global topology information of a static network. It is difficult for such methods to learn desirable features for the nodes in an incomplete graph whose topology information is sparse or the new nodes in a dynamic graph. It is also challenging for them to deeply incorporate node attributes as the complementary information to improve the network embedding performance. To this end, in this paper we propose a Multi-View Adversarial learning based Network Embedding model named MVANE to deeply fuse the network topology information and node attributes to better perform network embedding on incomplete graphs. The insight is that the network topology and the node attributes are treated as two correlated views. The learned embedding vector of a node should be able to reveal its unique characteristics in both views. Specifically, the adversarial autoencoder is introduced as the basic model of MVANE. Autoencoder can learn a projection function to directly map the input feature vectors into the latent space, which ensures the MVANE learn embeddings for the new nodes through features projection without the need of retraining the model. Meanwhile, the adversarial learning strategy is also applied to better capture the cross-view correlations. The idea is that the learned embeddings in one view can not only reconstruct the inputs in this view, but also generate the features in another view. Under a unified learning framework, the latent representations in different views are fused and jointly reinforced by the proposed self/cross-view learning model. Empirically, we evaluate MVANE over multiple network datasets, and the results demonstrate the superiority of our proposal. • We investigate the novel problem of network embedding on incomplete graphs. • We propose an adversarial learning based network embedding model MVANE. • The experimental results demonstrate the superiority of the proposed model. [ABSTRACT FROM AUTHOR]