Self-developed malware was usually used by advanced persistent threat (APT) attackers to launch APT attacks. Therefore, we can enhance the understanding and cognition of APT attacks by comprehending the behavior of APT malware. Unfortunately, the current research cannot effectively explain the relationship between the recognition, detection, and defense of APT. The model of similar studies also lacks an explanation about it. To defend against APT attacks and inquire about the similarity of different APT attacks, this study proposes an APT malware classification method based on a combination of multiple deep learning algorithms and transfer learning by collecting malware used in several famous APT groups in public. By extracting the application programming interface (API) system calls, with the vector representation of features by combining dynamic LSTM and attention algorithm, we can obtain API at different APT families classification contributions trained dynamic. Thus, we used transfer learning to perform multiple classifications of the APT family. This study aims to reduce the burden of network security staff from reviewing a large number of suspicious files when defending against APT attacks. Additionally, it can effectively intercept them in the initial invasion stage of APT to perform targeted defense against specific APT attacks by combining threat intelligence in public. The experimental result shows that the proposed method can achieve 99.2% in distinguishing common malware from APT malware and assign APT malware to different APT families with an accuracy of 95.5%.

Our work focuses on the exploration of the internal relationships of signals in an individual sensor. In particular, we address the problem of not being able to evaluate such intrasensor relationships due to missing rich and explicit feature representation. To solve this problem, we propose GraphSensor, a graph attention network, with a shared-weight convolution feature encoder to generate the signal segments and learn the internal relationships between them. Furthermore, we enrich the representation of the features by utilizing a multi-head approach when creating the internal relationship graph. Compared with traditional multi-head approaches, we propose a more efficient convolution-based multi-head mechanism, which only requires $56\%$ of model parameters compared with the best multi-head baseline as demonstrated in the experiments. Moreover, GraphSensor is capable of achieving state-of-the-art performance in the electroencephalography dataset and improving the accuracy by $13.8\%$ compared to the best baseline in an inertial measurement unit (IMU) dataset.

In the era of intelligent applications, Mobile Edge Computing (MEC) is emerging as a promising technology that provides abundant resources for mobile devices. However, establishing a direct connection to the MEC server is not always feasible for certain devices. This paper introduces a novel Device-to-Device (D2D)-assisted system to address this challenge. The system leverages idle helper devices to execute and offload tasks to the MEC server, thereby enhancing resource utilization and reducing offload time. To further minimize offloading time for latency-sensitive tasks, this paper incorporates edge caching. The problem is formulated by jointly optimizing computation, communication and caching, and a novel Joint Multiple Decision Optimization Algorithm (JMDOA) is proposed to solve the minimum-energy-consumption problem. Specifically, the JMDOA algorithm decomposes the integer-mixed non-convex optimization problem into two subproblems based on distinct properties of discrete variables. These subproblems are solved separately and optimized iteratively, ensuring convergence to a suboptimal solution. Simulations demonstrate the effectiveness and superiority of JMDOA, exhibiting lower energy consumption and reduced time compared to other baseline algorithms, approaching the optimum. This work contributes to the field by presenting a novel approach to optimizing resource allocation in MEC systems, with potential implications for the future development of intelligent applications.

Few-shot action recognition seeks to recognize novel actions with limited labeled examples. While dual-modal approaches incorporating video and textual modalities offer enhanced semantic context, existing methods often rely on naive feature fusion strategies, failing to capture deep semantic correlations across modalities and limiting generalization. We propose DPCA-Net, a dual-modal metric learning framework that constructs a unified dual-prototype consistency alignment space. DPCA-Net explicitly models distributional, structural, and metric consistency across modalities to enhance prototype quality and similarity estimation. It integrates three core components: (1) Frame-wise Text-guided Modeling (FTM), which uses conditional prompt learning to embed video frame-level visual features into the textual space, achieving structural consistency; (2) Dual-Modal Metric Learning via dual-path Dynamic Time Warping (Dual-DTW), jointly aligning visual and cross-modal prototypes to ensure metric consistency; and (3) Distribution Consistency Mapping (DCM), which leverages Maximum Mean Discrepancy and cosine similarity to align support-query distributions and reinforce representation robustness. Extensive experiments on three benchmark datasets show that DPCA-Net consistently outperforms prior methods. It surpasses CLIP-FSAR by 1.3%–2.7%, achieving 89.7% (1-shot) on Kinetics and 99.12% (5-shot) on UCF-101. These results highlight the effectiveness of consistency-driven prototype alignment for robust and generalizable cross-modal few-shot action recognition.

This work develops a generalizable neural network, SENSORNET, for sensor feature learning across various applications. The primary challenge addressed is the poor portability of pretrained neural networks to new applications with limited sensor data. To solve this challenge, we design SensorNet, which integrates the flexibility of self-attention with multi-scale feature locality of convolution. Moreover, we invent patch-wise self-attention with stacked multi-heads to enrich the sensor feature representation. SensorNet is generalizable to pervasive applications with any number of sensor inputs, and is much smaller than the state-of-the-art self-attention and convolution hybrid baseline ($0.83$ M vs. $3.87$ M parameters) with similar performance. The experimental results show that SensorNet is able to achieve state-of-the-art performance compared with the top five models on a competition activity recognition dataset (SHL’18). Moreover, pretrained SensorNet in a large inertial measurement unit (IMU) dataset can be fine-tuned to achieve the best accuracy on a much smaller IMU dataset (up to $5\%$ improvement in WISDM) and to achieve the state-of-the-art performance on an EEG dataset (SLEEP-EDF-20), showing the strong generalizability of our approach.