With the rapid evolution of artificial intelligence (AI) models, remote sensing applications increasingly rely on computationally intensive deep learning algorithms. However, the outdated onboard computing hardware and limited processing capacity of a single satellite can no longer meet these growing demands. While offloading data to the ground for processing is a common alternative, it faces significant challenges due to bandwidth-limited and unstable satellite-to-ground communication links. To overcome these limitations, we propose a flexible framework that leverages existing in-orbit satellite computing resources to accommodate AI models of varying scales and complexities. As a representative application scenario, this paper focuses on in-orbit object detection of remote sensing images. By aggregating idle computing resources from multiple satellites, the proposed framework enables distributed onboard image processing, which significantly reduces the volume of data that needs to be transmitted to the ground. To achieve this, we partition convolutional neural networks (CNN) models into multiple sequential slices and deploy them across satellites, taking into account model complexity, hardware heterogeneity, and dynamic inter-satellite links (ISLs). We formulate the model partitioning problem as a Series Multi-Armed Bandit (SMAB) problem and propose both online and offline algorithms to solve it. Compared to the traditional method of transmitting raw remote sensing data to ground stations via ISLs for post-processing, our approach reduces end-to-end latency by 35%.

Encrypted traffic classification plays an important role on network security and network management. However, existing methods face two key limitations: they require large volumes of labeled data and are typically designed under the closed-world assumption. This paper presents a novel approach combining pre-training techniques with contrast learning to address the challenges of open-world encrypted traffic classification. Our method utilizes a BERT-based pre-training model to learn generic traffic representations from large-scale unlabeled encrypted traffic. In this stage, two pre-training tasks, the Masked BURST Model (MBM) and BURST Context Prediction (BCP), are introduced to capture both structural and contextual features of the traffic. Following this, a contrast learning strategy is applied on a small labeled dataset to learn a discriminative feature space, where traffic from the same category is clustered closely, and traffic from different categories is largely separated. This strong discriminative property enables high-accuracy classification of known traffic categories while detecting novel ones. Experimental results on five encrypted traffic classification tasks demonstrate that: (1) under full data conditions, our approach outperforms state-of-the-art methods with up to an 11.60% improvement in F1 score in both closed- and open-world scenarios; (2) even when working in a few-shot mode, our method achieves the best performance in three out of five tasks in the closed-world setting and consistently outperforms the best methods by up to 9.99% in terms of F1 score in the open-world scenario.

Encrypted traffic classification requires capturing robust and effective traffic representations from data that lack explicit patterns and clear semantics, which is crucial for network management and cybersecurity. Existing methods heavily rely on large amounts of labeled data or expert-designed features and struggle to generalize across different classification scenarios. Leveraging unlabeled traffic data to learn universal representations of encrypted traffic remains a key challenge. In this paper, we propose a novel traffic representation model called ET-Former. ET-Former learns universal representations of various encrypted traffic from large-scale unlabeled data and can be fine-tuned with a small amount of labeled data for specific tasks. ET-Former achieves state-of-the-art performance in four of five encrypted traffic classification tasks, demonstrating exciting features such as robustness, generalization, and accuracy.

Large-scale low earth orbit (LEO) satellite constellations have become important to support ubiquitous communications. The vast computational resources embedded within satellite networks are driving the rapid development of space computing and satellite Cloud-native technologies, with various satellite artificial intelligence (AI) applications. However, a single satellite only has limited communication and computation resources, especially when dealing with remote sensing images by deep learning methods. To address this issue, this paper proposed a pipeline inference approach for Cloud-native satellites: PipeSIA, which enables multiple satellites to execute inference tasks cooperatively. A satellite constellation was designed for PipeSIA, whose network topology was also organized as a pipeline for efficient communication. Different from conventional computation offloading schemes, the proposed scheme jointly considers the node sequence and task allocation problem. A joint optimization problem is formulated as a series multi-arms bandit (SMAB). Then$\epsilon$-greedy algorithm and deep Q-network (DQN) are applied to optimize the node selection and task allocation. Simulation results show that the proposed scheme can effectively reduce the overall delay.

Middleboxes are becoming indispensable in modern networks. However, programming the network stack of middleboxes to support emerging transport protocols and flexible stack hierarchy is still a daunting task. To this end, we propose Rubik, a language that greatly facilitates the task of middlebox stack programming. Different from existing hand-written approaches, Rubik offers various high-level constructs for relieving the operators from dealing with massive native code, so that they can focus on specifying their processing intents. We show that using Rubik one can program the middlebox stack with minor effort, e.g., 250 lines of code for a complete TCP/IP stack, which is a reduction of 2 orders of magnitude compared to the hand-written versions. To maintain a high performance, we conduct extensive optimizations at the middle-and back-end of the compiler. Experiments show that the stacks generated by Rubik outperform the mature hand-written stacks by at least 30% in throughput.

Heterogeneous network (HetNet) has been proposed as a promising solution for handling the wireless traffic explosion in future fifth-generation (5G) system. In this paper, a joint subchannel and power allocation problem is formulated for HetNets to maximize the energy efficiency (EE). By decomposing the original problem into a classification subproblem and a regression subproblem, a convolutional neural network (CNN) based approach is developed to obtain the decisions on subchannel and power allocation with a much lower complexity than conventional iterative methods. Numerical results further demonstrate that the proposed CNN can achieve similar performance as the Exhaustive method, while needs only 6.76% of its CPU runtime.