Abstract:

Abstract:Convolutional Neural Networks (CNNs) as a quintessential representation of deep learning, are the most commonly used neural networks in tasks such as computer vision. However, convolution operations typically account for over 90% of the runtime in CNNs, becoming a bottleneck for performance. Additionally, due to the complexity of current hardware and the diversity of workloads, specific optimizations in previous work often lack performance portability. To address this problem, the author introduces BlazerML, an open-source convolution computation library based on auto-generated code templates from TVM, capable of automatically generating high-performance convolution implementations for any input shape. BlazerML is implemented based on the Winograd algorithm, known for its high performance in fast convolution algorithms. Experimental results demonstrate that BlazerML significantly outperforms current state-of-the-art open-source libraries. On x86 CPUs, running common deep learning network forward inferences, it is faster by 1.18—2.47 times, 1.18—2.27 times, and 1.01—1.66 times compared to OnnxRuntime, MNN, and the TVM community version, respectively. On ARM CPUs, for single-layer inference of common deep learning networks, it surpasses ACL and FastConv by 1.26—6.11 times and 1.04—4.28 times, respectively.

Abstract:Existing methods for video traffic identification are mainly targeted at specific platforms and mostly require capturing full flows, which makes them unsuitable for high-speed networks management. This paper proposes a method for video traffic identification from multi-platforms in the sampled high-speed traffic.This paper analyze multiple video platform transmission protocols, extract features based on their common characteristics to construct a composite feature space, and further process these features to eliminate the effect of sampling on feature stability. Then, feature vectors are extracted and a classification model is trained. In the experiments, high-speed networks traffic with a bandwidth of 10 Gbps and a sampling rate of 1:32 was used. The results showed that the proposed method can quickly identify video traffic from multi-platforms with a precision of over 98%.

Abstract:In response to the current power plant switch detection methods that are unable to cope with realworld open-set environments and the low accuracy in recognizing rare categories, the target recognition problem is transformed into a similarity measurement issue, and a new algorithm is proposed. The new algorithm is based on the triplet network of deep metric learning, using a ResNet-18 with an added SE Block to extract features, and enhances learning effects by cross-batch mining. To evaluate the performance of the algorithm, a dataset with 3 300 switch images was created. The algorithm was tested on the self-built dataset for closedset testing, open-set testing, and few-shot testing. The experimental results show that the algorithm demonstrates excellent discrimination ability in the closed-set state. It can not only accurately identify the categories in the training set but also effectively distinguish states that were not encountered during training and those with lower occurrence frequencies. This capability indicates that the algorithm is not only suitable for real-world open-set environments but also significantly improves the recognition accuracy for small-sample data.

Abstract:In the field of encrypted mobile application traffic classification, traditional methods classify traffic based on the characteristics of bidirectional traffic. However, in actual scenarios, asymmetric routing will cause remote network administrators to only obtain unidirectional traffic, which will reduce the accuracy of traditional methods. Therefore, an encrypted mobile application traffic classification method using only one-way traffic characteristics is designed. Since downlink traffic contains more information than uplink traffic, the payload of downlink traffic is chosen for analysis. Due to the temporal and spatial correlation of mobile application traffic, a bidirectional long short-term memory network is proposed to capture the temporal correlation of data streams, a convolutional neural network is used to learn the spatial correlation of features, and an attention layer is introduced to focus on important features to further improve the recognition accuracy. Compared with the previous methods, this method has a wider range of use, can be applied to both unidirectional and bidirectional traffic scenarios, and uses fewer features to obtain higher accuracy.

Abstract:In multi-party conversations, identifying the reply-to relation between messages is an important task in the dialogue domain. Existing efforts have not addressed the following two issues related to data distribution: shorter messages tend to appear more frequently, while shorter texts contain less semantic information, which limits the learning ability of the model; the number of positive samples with reply-to relation is often much less than the number of negative samples, leading to data skewness issue during training phase and reducing the model’s performance in processing positive samples. Aiming at the two issues, this paper proposes an improved model based on a pre-trained language model, which firstly mitigates the short text-related issue through dynamic inquiry window modeling; and then copes with the positive sample-related issue through position-driven positive sample weight optimization. The paper is compared with previous research, and the experimental results show that this paper’s work improves the recall metric by an average of 15.7% compared to the baseline model based on the pre-trained language model. In addition, this paper constructs a new dataset collected from the Telegram platform, which can provide data support for subsequent related studies.

Abstract:Metaphor has the purpose of inspiring understanding and persuading others. Currently, metaphor presents the trend of multimodal integration of text, images, and videos. Therefore, identifying the metaphorical semantics contained in multimodal contents has research value for Internet content security. Due to the lack of multimodal metaphor datasets, it is difficult to establish research models. Therefore, current scholars pay more attention to text-based metaphor detection. To overcome this shortcoming, the paper first generates a new multimodal metaphor dataset from the perspectives of image-text, metaphor appearance, emotion expression, and author intention. Then, Kappa scores were used to assess the consistency among the annotators of the dataset. Finally, a multimodal metaphor detection model is constructed to verify the quality and value of the multimodal data set by combining image attribute features, image entity features, and text features with the help of a pre-training model and attention mechanism. The experimental results show that the metaphor dataset with emotion and intention can improve the effectiveness of metaphor model detection, and confirm that the interrelationship of multimodal information is helpful for understanding metaphor.

Abstract:With the rapid development of Internet technology, network security issues have become increasingly prominent. Among these, the identification and classification of encrypted traffic have emerged as significant research directions. This paper provides a comprehensive review of current machine learning-based techniques for encrypted traffic classification. First, it briefly introduces common encryption protocols and their characteristics from a layered perspective. Then, it provides an overview of the datasets and evaluation metrics used in this field. Furthermore, a discussion on encrypted traffic analysis methods based on traditional machine learning and deep learning is conducted, with a detailed analysis of key techniques such as feature engineering and classifier models. Finally, it summarizes the challenges currently faced in this field, including the lack of interpretability and the risk of adversarial examples, and looks ahead to future research directions aimed at enhancing interpretability, automating feature extraction, and automating optimizing model structures.

Abstract:The unique financial characteristics make smart contract, which are widely nested in various blockchain platforms, one of the most successful applications of blockchain technology. Due to the high economic value of carrying large amounts of assets and virtual currencies, smart contracts are constantly subject to various security attacks. In addition, features such as anonymity and automatic execution make smart contracts used in a variety of illegal transactions and malicious applications. Based on this, the paper firstly introduces the operation mechanism and principle of smart contracts in blockchain-related technologies, discusses the application scenarios of smart contract technology and the potential security vulnerabilities and security problems that may be triggered by the development of smart contract technology, and then, based on the summary of the existing research work, discusses the challenges faced by the field of smart contract vulnerability detection, and looks forward to the future research direction of smart contracts in combination with deep learning technology.