Yu Guo, Shenghao Ye, Shuangwu Chen, Zijian Wen, Tao Zhang, Bai Qirui, Dong Jin, Yunpeng Hou, Huasen He, Jianyang, Xiaobin Tan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zijian Wen, Tao Zhang, Shuangwu Chen, Shenghao Ye, Yu Guo, Qirui Chen, Jingxian Shuai, Yunpeng Hou, Huasen He, Jianyang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shenghao Ye, Yu Guo, Dong Jin, Yuxiang Wang, Yikai Shen, Yunpeng Hou, Shuangwu Chen, Jianyang, Xiaofeng Jiang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Motivated by the intelligent decision-making ability, Deep Reinforcement Learning (DRL) has been extensively applied in multi-Unmanned Aerial Vehicle (UAV) trajectory planning. This article investigates the application of DRL in Three-Dimensional (3D) trajectory planning in environments with obstacles, where multiple UAVs act as aerial Base Stations (BSs) to provide services to ground user hotspots. The existing DRL-based algorithms require a significant amount of trial and error iterations to obtain sufficiently high-performing agents. To cope with this drawback, we introduce Transfer Learning (TL) into DRL, enabling the UAV agent to possess prior knowledge upon initialization, thereby quickly adapting to unfamiliar environments and significantly improving performance. Considering the limited local observations of UAVs, a multi-modal fusion autoencoder is proposed for extracting and compressing cross-modal features from global observations to obtain the global fusion state, which enhances the perception capabilities of UAVs without incurring excessive communication overhead. Finally, we propose an exploration novelty-driven collaborative trajectory planning algorithm for multiple UAVs, which ensures obstacle avoidance and enhances the UAVs’ exploration capabilities to cover all hotspots. We adopt a probabilistic channel model and discretize both the time and DRL action space to achieve a balance between practicality and tractability. Extensive experiments demonstrate that our proposed transfer reinforcement learning method can improve the initial performance of the UAV agents by 76%. The perception and exploration-enhanced trajectory planning algorithm significantly increased the exploration efficiency and improved hotspot coverage by 50%.

Logical anomalies occur when a product's assembly violates prescribed logical rules, which widely exist in industrial assembly and packaging processes. Due to the difficulty in comprehending such complex logical relationships, a paucity of research has focused on the industrial logical anomaly diagnosis (LAD). Recently, large language models (LLMs) have demonstrated strong semantic understanding and zero-shot reasoning capabilities, making them a promising tool for LAD. However, directly applying LLMs to LAD still face two critical challenges: 1) the scarcity of abnormal samples in real-world settings, and 2) the propensity of LLMs to generate hallucinated or unreliable diagnostic conclusions. To address these challenges, we propose LogiDiag, a novel diagnostic reasoning method for LAD, to pinpoint where and why a product fails to comply with the logical rules, thereby elevating product quality and reducing remedial intervention cost. We design a visual descriptor that identifies product component attributes even in out-of-distribution abnormal images and organizes them into component descriptions for LLMs' comprehension. To mitigate hallucinations, we devise a planner to guide LLMs in diagnostic reasoning through rule orchestration, tool allocation, and chain-of-diagnosis generation. Experimental results on multiple benchmark datasets validate the competitive performance of LogiDiag.

The exponential surge in network traffic has imposed significant challenges on traditional Internet architectures, resulting in high latency and redundant transmissions. Cache-enabled networks alleviate these issues by deploying content closer to end-users, making the planning of such networks a research focus. However, regional heterogeneity in user demand and caching interdependencies among hierarchical nodes complicate the planning process. Most existing approaches rely on simplistic even allocation or empirical methods, which fail to simultaneously meet user performance expectations and minimize deployment costs. This paper proposes a network planning framework based on the Equivalent Characteristic Time Approximation (ECTA). The approach begins by establishing a performance–resource mapping. Using ECTA, we decouple the tightly coupled characteristic time relationships across hierarchical nodes, thereby accurately estimating the required cache capacity and bandwidth needed to achieve user performance targets. Building on this foundation, we formulated the network planning as a constrained convex optimization problem that minimizes deployment cost while satisfying user performance constraints. We conducted extensive experiments on a large-scale simulation platform (ndnSIM) and a real-world cache-enabled network testbed (CENI-HeFei). The results demonstrate that, under identical network topologies and total resource constraints, our method significantly improves cache hit probability while reducing deployment costs compared to homogeneous resource allocation schemes. This work provides a practical theoretical foundation and valuable insights for the design, deployment, and optimization of future cache-enabled networks.

The massive parameter scale of sparsely-activated Mixture-of-Experts (MoE) models necessitates distributed training with hybrid parallelism. Placing such training tasks,i.e.mapping the logical partitions of an MoE model to available physical NPUs, is challenging. Due to the bandwidth and latency discrepancies between intra- and inter- Pods, the cross-Pod communication usually becomes a bottleneck. The high dispersion of NPUs in multi-tenant clusters exacerbates this issue further. However, a paucity of studies has considered the cross-Pod model placement problem. To address this challenge, we propose a novel model placement scheme tailored for MoE model training with hybrid parallelism in multi-tenant clusters. By quantifying the cross-Pod communication overhead incurred during MoE model training, the model placement is formulated as a 0-1 integer quadratic problem, which is NP hard. Motivated by the traffic difference between different parallelism, we decompose this problem into two subproblems. To solve the subproblems, we propose a lightweight two-stage algorithm based on Best-Fit strategy and neighborhood search. Experiments under different models and network topologies show that our model placement scheme can reduce cross-Pod traffic by 35.9% and cut communication time by 18.7% compared to state-of-the-art methods.

The surge of artificial intelligence (AI) has intensified compute-intensive tasks, sharply increasing the need for energy-efficient management in geo-distributed data centers. Existing approaches struggle to coordinate task scheduling and cooling control due to mismatched time constants, stochastic Information Technology (IT) workloads, variable renewable energy, and fluctuating electricity prices. To address these challenges, we propose D3T, a dual-timescale deep reinforcement learning (DRL) framework that jointly optimizes task scheduling and thermal management for energy-efficient geo-distributed data centers. At the fast timescale, D3T employs Deep Q-Network (DQN) to schedule tasks, reducing operational expenditure (OPEX) and task sojourn time. At the slow timescale, a QMIX-based multi-agent DRL method regulates cooling across distributed data centers by dynamically adjusting airflow rates, thereby preventing hotspots and reducing energy waste. Extensive experiments were conducted using TRNSYS with real-world traces, and the results demonstrate that, compared to baseline algorithms, D3T reduces OPEX by 13% in IT subsystems and 29% in cooling subsystems, improves power usage effectiveness (PUE) by 7%, and maintains more stable thermal safety across geo-distributed data centers.

Fault detection (FD) is essential for wafer quality control in semiconductor manufacturing (SM), as it can identify abnormal wafers in the early stages. However, frequent chamber discrepancies leads to distribution shifts in the data from different chambers, which causes performance degradation of the existing model. In this paper, we conceive a robust cross-chamber fault detection method in SM, which to the best of our knowledge is the first work that employs domain generalization to address the issue of cross-chamber fault detection in SM. Our basic idea is to map the samples from the source chambers and target chambers to the same hypersphere space, making normal samples cluster around a specific feature center while faulty samples stay away from it. Due to the scarcity of faulty samples, we propose a one-class classification-based fault detection method relying on normal samples to establish classification boundaries. To generalize the model to unseen target chambers, we design a meta-learning-based one-class domain generalization approach. We also devise a strategy to enhance distribution alignment within hypersphere, making the classification boundaries of various chambers be close to each other. The evaluations on real-world data collected from a wet process equipment in SM verify the high robustness of the model to chamber discrepancies with limited faulty samples.

The integrated merits of Unmanned Aerial Vehicle (UAV) networks including high mobility, ease of deployment and low cost have promoted their widely application in both civilian and military areas. However, the complex communication environments, dynamic network topology and intermittent links pose significant challenges to the transmission reliability of UAV networks. Existing end-to-end reliable transmission mechanisms rely on feedbacks from receivers to trigger retransmission, which impose extra transmission delay and redundant retransmission. In this work, we propose a Hop-by-Hop Redundancy-guaranteed Adaptive Coding (HHRAC) approach for enhancing transmission reliability of dynamic UAV Networks. To cope with lossy links, a link quality-adaptive coding algorithm is proposed, which dynamically adjusts coding redundancy rate according to link quality. Meanwhile, the Cauchy matrix is employed to design efficient coding matrices, which greatly improve decoding efficiency and enable the intermediate nodes to perform low-complexity verification. Moreover, a redundancy-guaranteed hop-by-hop transmission mechanism is provided to avoid End-to-End (E2E) retransmission and ensure the destination node has a high probability to receive sufficient packets. To avoid receive queue overflow, we further propose a queue length prediction based congestion control algorithm to control the sending rates of UAVs. The experimental results show that HHRAC achieves significant performance gains compared to existing algorithms in terms of transmission delay and retransmission times.

Space-Ground Integrated Networks (SGINs) hold the potential to enable seamless and high-quality global coverage in an economically viable manner. However, the limited bandwidth and relatively long delay of satellite-ground links pose significant challenges in meeting the increasing demands driven by surging traffic. To address this issue, we propose a three-tier cooperative caching architecture incorporating base stations (BSs), satellites, and a content server, which aims to reduce the average content retrieval delay through cooperative caching between BSs and satellites. The joint optimization problem among them is both challenging to solve and non-scalable due to its exponentially increasing computational complexity. Meanwhile, the geographic characteristics of requests are overlooked in existing work. As a consequence, we propose a novel block partitioning algorithm based on the popularity similarity across areas to facilitate cooperative caching, which reduces the computational complexity, and ensures the scalability and the effectiveness of cooperative caching. Based on the partitioned blocks, the optimization problem is decomposed into two subproblems: intra-block cooperative optimization among BSs and inter-block cooperative optimization among satellites, effectively catering to the characteristics of wide-area coverage in SGINs. For the subproblems with finite dimensions, Semidefinite Relaxation (SDR) based intra-block and inter-block cooperative caching approaches are proposed to obtain the optimal cooperative caching strategies. Extensive simulations demonstrate that, compared to directly solving the original problem, the proposed algorithm reduces the solving time from exponential to linear growth. Moreover, our algorithm outperforms existing schemes by reducing 13% average retrieval delay and improving 12% overall cache hit rate.

Low Earth Orbit Satellite Networks (LEOSNs) have emerged as a promising paradigm for space information networks, where multiple inter-satellite links facilitate the rapid transmission of on-orbit data to multi-ground station systems. When the destination of the transmission is not a certain ground station, but anyone of the ground stations, it can be modeled as an anycast problem. However, the time-varying topology, dynamic inter-satellite link status and limited onboard resources bring challenges to the routing of on-orbit data. Existing unicast routing solutions failed to address the anycast routing problem as they could not fully utilize multiple ground stations. Inspired by the Potential Field (PF) theory in physics, we make the first attempt to adopt the PF approach in the on-orbit data anycast routing problem. We design a synthesized PF model including several sub-fields corresponding to network states such as length of path, node transmission load and link bandwidth. By implementing inter-satellite propagation and synthesis of PF, dynamic perception and unified measurement of network states can be achieved. Based on our PF model, we propose a distributed PF-based and Network State-aware Anycast Routing (PFNSAR) algorithm, which regards the ground stations as multiple sources of attractive potential and guides packets along the gradient of synthesized PF. Meanwhile, the occurrence of the well-known local minimum is novelly handled by setting PF configuration subject to a parameter constraint and switching routing modes when routing packets. Extensive simulations demonstrate that PFNSAR provides unified assessment of multi-dimensional network states, reduces up to 30% of average delivery delay, and improves the performance including packet delivery rate and load balancing compared with existing works.

Website Fingerprinting (WF) attacks have posed a serious threat to the anonymity of the onion router (Tor) communication system, as attackers can passively pry into the encrypted traffic and infer the website visited by users. To defend against WF, recent studies focus on adversarial perturbations. However, most of them suffer from a high bandwidth overhead and a low defense performance. To address this problem, our basic idea is to generate perturbation only on the sensitive regions, which can effectively mask the website’s fingerprint, thus misleading the WF attack models and reducing the bandwidth overhead. In this paper, we formulate a joint optimization problem of perturbation position and magnitude by confining the perturbations within sensitive regions, which is rarely considered in the literature. We propose a robust low-overhead WF defense method based on reinforcement learning (RL), named RLpatch. RLpatch identifies the common sensitive regions of various surrogate models and adjusts perturbation according to the query result from a query WF model. It further employs the positional frequency of perturbations to generate a common perturbation paradigm for different traces of a same website. Experimental results show that RLpatch achieves higher defense performance, lower bandwidth overhead and better robustness against adversarial training compared to the state-of-the-art methods.

Robust Privacy-Preserving Federated Learning (PPFL) is a secure distributed machine learning paradigm designed for untrusted environments, aiming to aggregate gradients while ensuring the reliability of the results without disclosing user gradients. However, existing single-server robust PPFL schemes require users to generate commitments in each aggregation round to ensure the correctness of the aggregation results, which leads to high computational overhead. We propose an efficient single-server robust PPFL scheme named Privacy-Preserving Robust Federated Learning (PRFL). PRFL achieves efficient gradient aggregation through a “detection-identification-exclusion” strategy. PRFL only performs quick detection without requiring high-complexity commitments in most of aggregation round, thereby ensuring excellent efficiency. PRFL comprises three pivotal components: Privacy-Preserving Gradient Aggregation Based on Packed Secret Sharing (PGAPS), Swift Share Verification based on Dual Codes (SSVDC), and Commitment-based Malicious User Identification (CMUI). PGAPS is utilized to implement the FLTrust rule without disclosing gradients. SSVDC swiftly detects incorrect shares without using commitments. CMUI identifies malicious users when SSVDC detects incorrect shares. Experimental results demonstrate the robustness and efficiency of PRFL. In a PPFL system with 100 users, PRFL can robustly aggregate gradients of a million dimensions within 37 seconds of average computational time.

Microservice architecture offers great scalability and flexibility to the development of online services systems. Performance anomalies, which happen frequently due to code bugs or runtime environment misconfiguration, can severely damage the system availability and cause great losses. However, it is challenging to detect performance anomalies and locate their root causes considering the large volume of monitoring data (e.g., metrics and traces) and the complex dynamic interdependence between heterogeneous services. Against these challenges, we propose HG-PAD, an automatic performance anomaly diagnosis (PAD) framework for microservice systems. We build the multi-relation heterogeneous graph to model the intricate dependency between services. We further design a structure learning mechanism combining graph neural network (GNN) and node embedding learning to capture the dynamic and latent dependencies. Based on the optimized dependency graph, we devise a Conditional Variational Auto-Encoder (CVAE) based unsupervised anomaly detection method and a graph attention network (GAT) based root cause localization method for accurate anomaly diagnosis. We use datasets of different scales based on real world applications to verify the effectiveness of HG-PAD, and the experimental results show that HG-PAD achieves better diagnostic performance compared with existing baseline methods.

Encrypted tunnels are increasingly applied to privacy protection, however, a passive eavesdropper can still infer which website a user is visiting via website fingerprinting (WF). State-of-the-art WF suffers from several critical challenges in a realistic multi-tab web browsing scenario, where the number of concurrent tabs is dynamic and uncertain, training a separate model for each website is too overweight to deploy, and the robustness against the packet loss, duplication and disorder caused by dynamic network conditions is rarely considered. To address these challenges, we propose a robust and lightweight multi-tab WF method over the encrypted tunnel, named RobustWF. Due to the causality relationship between user’s request and website’s response, RobustWF employs causality correlation to associate the interactive packets belonging to the same website together, which form a causality chain. Then, RobustWF utilizes context learning to capture the dependencies between the causality chains. The missing of some specific details does not have a significant impact on the overall structure of target web, thus enhancing the robustness of RobustWF. To make the model lightweight enough, RobustWF trains an integrated model to adapt to the dynamic number of concurrent tabs. The experimental results demonstrate that the accuracy of RobustWF improves 14% in dynamic multi-tab WF scenario compared to the State-of-the-art method.

The Low Earth Orbit (LEO) satellite constellation has been recognized as an important component of the future 6G network. However, due to the high-speed movement of LEO satellites and the potential for link failures, achieving optimal satellite communication performance with a static inter-satellite links (ISLs) scheme is challenging. To solve this problem, this paper proposes an ISL re-planning algorithm with considering link failures based on multi-agent deep reinforcement learning (named ReISL). In ReISL, a multi-objective optimization problem is formulated to maximize the system capacity while minimizing the link switching costs. Then, multi-agent deep reinforcement learning is employed to derive the optimal ISL re-planning schemes, where each satellite utilizes Double Deep Q-Network (DDQN). Finally, extensive experiments are carried out and the results demonstrate that our proposed algorithm ReISL can outperform the baseline algorithms.

The deployment of UAVs as aerial base stations (BSs) has been considered as a promising supplement to the ground networks, which can quickly build an emergency communication network in a disaster area or significantly relief the communication burden imposed by hot-spots. However, the application of UAVs as aerial BSs is constrained by the limited onboard energy and communication coverage of UAVs. In particular, for a large target area, multiple UAVs should be deployed to meet the communication requirements. Therefore, designing the optimal trajectories of multiple UAVs is crucial to boost the UAV network performance. Inspired by the promising future of UAV BSs, this paper aims at proposing a distributed 3-dimensional (3D) trajectory design algorithm for multiple UAVs to optimize the system performance. We formulate the trajectory design problem as a multi-objective optimization problem to improve the user equipment (UE) access rate, ensure fair access opportunities, increase transmitted data volume and reduce energy consumption. Further inspired by the decision-making ability of deep reinforcement learning (DRL) in complex environments, we propose a DRL based trajectory design algorithm for multiple UAVs, namely DMTD, in which UAVs can explore both the optimal flight altitude and the potential UE distribution area in the iterative interactions with the environment, and then select the optimal flight trajectories to boost the network performance from multiple aspects. Extensive experimental results under different UE distributions have demonstrated that the proposed DMTD algorithm can find the optimal altitude to provide maximum coverage. Moreover, DMTD beats existing algorithms by providing high UE access rate, ensuring fair network service and increasing total transmitted data volume at the cost of a relatively low energy consumption. Especially in the scenes with dense and randomly distributed UEs, DMTD provides a UE access rate close to 0.9 and transmits 6 times of data volume than existing algorithms.

Low Earth Orbit Satellite Networks (LEO-SNs) have emerged as a promising paradigm for future space information networks. However, the time-varying topology, link intermittency, limited onboard resource and relatively long transmission distance imposed unprecedented challenges on guaranteeing the delay Quality of Service (QoS). In contrast with existing routing-based or resource optimization-based solutions, onboard processing provides an alternative way to reduce the transmission delay by dwindling the transmitted data size. The employment of onboard processing makes it critical to select a routing path with sufficient energy and properly allocate resources for transmission and processing. This paper studies the untouched onboard processing aided transmission delay minimization problem of LEO-SNs. A Distributed Network State Learning (DNSL) mechanism is proposed for synchronizing the network states, which induces Potential Field (PF) to model both the attraction of resources and the repulsion of transmission load. By jointly considering the channel conditions, onboard resources and transmission load, a Deep Q-network (DQN) based Intelligent In-orbit Routing (DIIR) algorithm is proposed for selecting a routing path with good channel conditions, sufficient energy and low transmission load to facilitate onboard processing. Moreover, a Deep Deterministic Policy Gradient (DDPG) based Intelligent Resource Allocation (DIRA) algorithm is provided to achieve intelligent and continuous resource allocation for exploiting onboard processing to minimize the transmission delay, while the resource and load states of satellites on the routing path are taken into consideration by including PF as an input. Extensive simulation results demonstrate that employing onboard processing with the proposed DIIR and DIRA algorithms significantly reduces the average transmission delay and packet loss rate.

Anomaly diagnosis relying on system logs to record runtime events is essential for improving the service quality of distributed systems and reducing economic losses. However, most existing log-based anomaly detection approaches depend on the assumption of the fixed quantitative or sequential patterns of a normal log event sequence. This assumption is challenged in the context of practical distributed and parallel systems due to the dynamic pattern of the log sequence, log data noise as well as concurrency of multiple anomalies. Against these challenges, this paper aims to perform the robust log-based anomaly diagnosis by capturing the event context information of the log event graph, called LogGraph, instead of straightforwardly employing the fixed quantitative or sequential patterns of the log records, thus reducing its sensitivity to the log flaws and the concurrency of multiple anomalies. Specifically, in order to handle multiple anomalies concurrency, LogGraph invokes the association rule to decouple log sequences. It further reinterprets a log record sequence into a log event graph modeled by event semantic embedding and event adjacency matrix. An attention-based Gated Graph Neural Network (GGNN) model is developed to capture the semantic information of the log graph, which enables the fine-grained and robust anomaly identification of the proposed scheme. We use real log data sets collected from Hadoop systems and network switches to verify the effectiveness of the proposed LogGraph in log data scenarios that contain noise and multiple anomalies concurrency problems. The experimental results show that the proposed LogGraph achieves high performance and strong robustness in anomaly diagnosis.

Website fingerprinting has emerged as a prominent topic in the area of network management. However, the proliferation of encrypted network traffic poses new challenges for website fingerprinting. In this paper, we analyze the behavior and correlations among the network flows generated by browsing a webpage and conclude that there exist specific spatio-temporal correlations among these network flows. Based on this finding, we propose the construction of an inter-flow spatio-temporal correlation graph (STCG) to model these correlations. In the STCG, each node represents a flow, with its features capturing the properties of the flow itself, and each edge with a weight vector represents the spatio-temporal correlation between two flows. Subsequently, we propose a graph neural network-based website fingerprinting method (STC-WF) by considering the inter-flow spatio-temporal correlations, in which the Graph Attention Network (GAT) and Self-Attention Graph Pooling (SAGPool) mechanisms are employed to acquire a comprehensive representation of the STCG. To evaluate the performance of STC-WF, we construct a real-world traffic dataset and conduct comprehensive evaluations. The experimental results demonstrate that STC-WF outperforms state-of-the-art methods in terms of accuracy and time consumption.

The proliferation of 360-degree video applications has brought significant challenges to existing networks. To meet the requirements of high transmission rate, low interaction latency, and high reliability, Mobile Edge Computing (MEC) has emerged as a promising technology that enables caching and processing at network edges. In this article, we present DACOD360, a deadline-aware content delivery system for the 360-degree video streaming over MEC networks. To address the challenges such as unpredictable viewports, uneven cached tiles, concurrent requests, and dynamic bandwidth, we formulate the deadline-aware delivery problem as a long-term integer program model to maximize the Quality of Experience (QoE) under the constraints of network bandwidth, cache capacity, and deadline. This optimization problem is a complex sequential decision that considers both deadline-constrained service quality at the temporal scale and multi-user resource allocation at the spatial scale. To solve it, we decompose the original problem into two sub-problems and solve them iteratively using Deep Reinforcement Learning (DRL) and Cooperative Bargaining Game (CBG). Comprehensive experiments are conducted in a wide variety of environments, and the results demonstrate that our proposed scheme outperforms the state-of-the-art schemes in terms of long-term QoE, traffic reduction, and other metrics.

Due to the concentrated distribution of network traffic, the Internet is highly vulnerable to link flooding attack in Distributed Denial-of-Service attacks (DDoS-LFA), which utilizes the legitimate low-rate attack traffic to block the selected network area. In recent years, building trusted networks has been considered as a promising strategy to address the security issues. Nevertheless, deploying a trusted link defense mechanism in the attacked network area faces many challenges imposed by the smart scheme and legitimate disguise of DDoS-LFA. In order to overcome these challenges, we propose a blockchain-based DDoS-LFA detection and mitigation scheme, named CREDIT, to guarantee the security of attacked area, while existing works only use blockchain to share the detection results of traditional solutions. CREDIT uses blockchain to record and share the information of links and flows in real time, which enables routers in the protected area to easily trace the paths of all active flows and capture the fragile links. On the basis of link features, a credible deep learning method performed on randomly selected nodes is proposed to detect DDoS-LFA against data spoofing. When an attack alarm is raised, CREDIT performs similarity analysis to locate attackers and migrate suspicious traffic based on the flow features of alarm links. Experimental results based on real implementation and attack testbed show that, by integrating blockchain, CREDIT performs better than traditional non-blockchain-based DDoS-LFA defense methods when faced with data tampering.

As a new computing paradigm, multi-data center computing enables service providers to deploy their applications close to the users. However, due to the spatio-temporal changes in workloads, it is challenging to coordinate multiple distributed data centers to provide high-quality services while reducing service operation costs. To address this challenge, this article studies the joint optimization problem of task scheduling and resource scaling in multi-data center systems. Since the task scheduling and the resource scaling are usually performed in different timescales, we decompose the joint optimization problem into two sub-problems and propose a two-timescale optimization framework. The short-timescale task scheduling can promptly relieve the bursty arrivals of computing tasks, and the long-timescale resource scaling can adapt well to the long-term changes in workloads. To address the distributed optimization problem, we propose a two-timescale multi-agent deep reinforcement learning algorithm. In order to characterize the graph-structured states of connected data centers, we develop a directed graph convolutional network based global state representation model. The evaluation indicates that the proposed algorithm is able to reduce both the task makespan and the task timeout while maintaining a reasonable cost.

Deploying microservices on the cooperative edge computing system greatly shortens the interaction delay between users and service and alleviates the traffic burden on the backbones, which has emerged as a new paradigm for service provision. However, it is challenging to embed microservices, having diverse resource demands and heterogeneous invocation relationships, into a distributed edge computing system with irregular network topology. In order to characterize the invocation relationship, we conceive a graph attention network based model to capture the structural features of microservices. Similarly, we propose a multi-channel directed graph convolutional network model to capture the spatial dynamic of edge resources distribution, which jointly considers the heterogeneity of the edge nodes and the links between them. Then, we develop a sequence-to-sequence based multi-step decision model, which maps the feature sequence of the current state to a sequence of deployment actions. Using this model, we further propose a microservice deployment algorithm based on graph neural network aided deep reinforcement learning, where a parallel asynchronous training process is used to accelerate convergence. The performance evaluation shows that the proposed algorithm can improve the deployment success ratio and resource utilization, while ensuring the load balance of edge nodes.

High-quality video can provide viewers with a better visual and immersive experience, but it typically requires a significantly higher network bandwidth to accommodate the larger amount of video data. The existing and commonly-used adaptive bitrate (ABR) approaches cannot provide high-quality video services for viewers when the network is poor. To address this issue, we propose an edge super-resolution assisted adaptive video streaming, named SuperABR, to improve viewers' Quality of Experience (QoE) and mitigate the influence of poor networks. First, to prevent the mismatch between the available computing resources and the VSR workload, the dynamics and trends of the available computing capability are extracted from a series of historical VSR reconstructing delays. Second, to optimize the video quality, rebuffering time, and video quality jitter for viewers, we formulate the optimization model by fully considering the states of all caching queues in SuperAbr and imposing probability constraints on the playback queue. Third, a queue-learning-based adaptive video streaming algorithm is proposed in SuperABR to jointly determine the source transmission resolution and the edge VSR-reconstructed resolution, which is essentially a Deep Reinforcement Learning (DRL) method with queue constraints. Finally, extensive experiments illustrate that SuperABR can improve QoE by 20%-76% compared to four baseline algorithms.

The IPv6 over low-power wireless personal area network (6LoWPAN) protocol stack is a promising solution to connect wireless sensor networks (WSNs) with the Internet for realizing a ubiquitous network interconnection of all things. However, 6LoWPAN networks face a critical challenge to control congestion caused by the burst of data traffic from wireless sensors. Packet loss will occur when the buffer overflows. This article focuses on the loss-tolerant congestion control problem in 6LoWPAN networks, which has not been addressed in existing works. We formulate the congestion control problem as a noncooperative Markov game framework and conceive a novel congestion control method, namely, deep reinforcement learning-aided loss-tolerant congestion control (DLCC), to alleviate congestion while maintaining a tolerable packet loss imposed by the buffer overflow. The proposed DLCC employs deep reinforcement learning (DRL) to solve the curse of state dimensionality, while packet loss constraints are handled by utilizing Lagrange multipliers to integrate the reward with loss constraints. By dynamically updating Lagrange multipliers in an online learning procedure, DLCC finds the optimal congestion control policy. Our simulation results show that DLCC maintains the packet loss rate below the tolerable threshold in the presence of congestion. In contrast to existing hybrid congestion control algorithms, the proposed DLCC algorithm is more energy efficient and provides higher throughput, lower average delay, and better fairness.

Widespread deployment of Small Satellite Networks (SSN) fosters the foreseen integration of space-air-ground networks to provide worldwide Internet access to oceanic and remote airspace. However, the dynamic topology of SSN, the lossy wireless link, and the limited transmission resources induce unprecedented challenges to Deterministic Satellite Network Transmission (DSNT) for the sake of improving the utility of the SSN facility. Motivated by these challenges, this work aims to develop a Deterministic Satellite Network Transmission approach with deterministic Spatio-temporal routing, Redundant coding and Multipath scheduling (DSNT-SRM) for bolstering superior communications of SSN. DSNT-SRM uses the ephemeris information and dynamic resource update mechanism to predict all upcoming communication opportunities and construct the deterministic spatio-temporal routing paths. By combining sparse and redundant network coding mechanisms, DSNT-SRM no longer cares about the arrival of each packet, but the number of coded packets it receives, since the lost packets can be compensated with deterministic redundant traffic. Moreover, the adaptive traffic balance between multiple spatio-temporal paths is designed to provide a deterministic delay guarantee when facing limited node resources and multi-user competition. Extensive experiments show that DSNT-SRM can achieve satisfactory performance improvement in reducing delay and improving delivery rate.

Self-piercing riveting (SPR) is widely used for joining lightweight and dissimilar materials in automotive body manufacturing, the quality of which directly affects the safety of vehicles. However, there is still no reliable method that can be used for SPR quality control without destructive test and manual intervention. This article presents an online nondestructive SPR defect detection method based on deep learning. By learning the temporal dependencies of punch force varying with rivet displacement under different joint combinations, the proposed method can provide real-time defect alarms and avoid the enormous cost of joint dissection. We develop an SPR parameter selection mechanism to rule out the irrelevant parameters, which enhances the learning performance. For the problem of model overfitting caused by the savage imbalance of SPR data, we design a conditional generative adversarial network based data generation model. In order to accommodate the difference in defect patterns between factory and laboratory, we devise a transfer learning based model migration method, which substantially reduces the amount of labeled factory data for model training. The evaluations on real SPR data collected from two car assembly lines of Audi and NIO verify that the proposed method achieves a high detection accuracy and a low missing rate in SPR defect detection.

The Tucker Decomposition (TKD) is able to provide the low-dimensional and informative representations of real-world large-scale tensorial data, which are necessary to extract potential features and enhance the original data. However, computing such decomposition directly for a dense tensor is usually computationally elusive, due to the repetitive operations of computing large-scale tensor-matrix product. Instead of direct decomposition, this paper proposes an efficient algorithm for seeking the Faster TKD of the large-scale tensor, which is a lightweight decomposition approach based on the technique of randomized sampling. The proposed algorithm first converts the original large-scale tensor into a small-scale subtensor via full-mode sampling operation, and then the core tensor of TKD can be computed directly based on the subtensor with low complexity. Finally, an approximate TKD of the original large-scale tensor can be obtained after sequentially computing approximate full-mode factor matrices. A theoretical error analysis is provided to show that the approximation error approximates zero with high probability, and the proposed algorithm is verified based on real tensorial data of$\text{23821.24}~GB$.

Driven by the prevalence of the computation-intensive and delay-intensive mobile applications, Mobile Edge Computing (MEC) is emerging as a promising solution. Traditional task offloading methods usually rely on centralized decision making, which inevitably involves a high computational complexity and a large state space. However, the MEC is a typical distributed system, where the edge servers are geographically separated, and independently perform the computing tasks. This fact inspires us to conceive a distributed cooperative task offloading system, where each edge server makes its own decision on how to allocate local computing resources and how to migrate tasks among the edge servers. To characterize diverse task requirements, we divide the arrival tasks into different priorities according to the tolerance time, which enables to dynamically schedule the local computing resources for reducing the task timeout. In order to coordinate the independent decision makings of geographically separate edge servers, we propose a priority driven cooperative task offloading algorithm based on multi-agent deep reinforcement learning, where the decision making of each edge server not only depends on its own state but also on the shared global information. We further develop a Variational Recurrent Neural Network (VRNN) based global state sharing model which significantly reduces the communication overhead among edge servers. The performance evaluation conducted on a movement trajectories dataset of mobile devices verifies that the proposed algorithm can reduce the task consumption time and improve the edge computing resources utilization.

Video services such as short video sharing have exploded due to the rapid development of Internet social media platforms. Caching video segments on satellites effectively shortens service delay and speeds up video sharing, especially for users without terrestrial Internet access. However, where to place what video and how to replace it in time is by no means an easy task, requiring careful consideration of many factors, e.g., satellite coverage, video popularity, and limited caching resource. In this paper, we propose a coverage-aware cooperative video caching algorithm (CACVC) that considers the prevalence of video in the coverage area and the collaboration between adjacent satellites. In CACVC, we model the cache placement problem of video as a Partially Observable Markov Decision Process (POMDP) to optimize the service delay of video provided by access satellites, neighboring satellites, or ground stations. We derive the optimal cache strategy by utilizing Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm with a centralized training and distributed execution paradigm. Simulation results show that the cache hit ratio can be improved by 4%~18%, and the average service delay can be reduced by 1%~14%.

Aeronautical Ad-hoc Networks (AANETs) have been proposed as the promising complement to terrestrial networks for promoting the global interconnection to provide in-flight network service, emergency communication, vessel traffic service, etc. However, the large network scale of AANETs may induce severe synchronization overhead when the traditional topology-based networking algorithms are adopted. Moreover, the high-dynamic topology and changeable flight routes make the existing position-based routing algorithms suffer loop routing and forwarding failure. Motivated by these problems, this paper aims at developing efficient and low-cost networking and routing algorithms relying on the concept of dynamic virtual topology which organizes the disordered topology of AANETs into a structural and simplified one. The basic idea is that each connected aircraft is assigned with a unique and sequentially increased Virtual Identifier (VID) and thus all the connected aircrafts are organized into a virtual cluster consisting of one trunk and several branches. An event-driven synchronization mechanism is leveraged for maintaining the virtual topology as well as relieving the communication burden imposed by periodical broadcasting. By jointly considering the geographic locations and the virtual locations of aircrafts, we formulate the routing problem in AANETs as a weighted distance minimization problem, and further propose a novel routing algorithm, namely Trunk-Branch Cooperation aided Routing (TBCR) algorithm. Specifically, TBCR employs the geographic greedy forwarding strategy for enhancing its flexibility and boosts the routing efficiency by adopting the loop-free virtual topology based local forwarding. For extending the networking and routing algorithms to the global range, a multi-domain routing solution is also provided. Extensive experimental results show that the proposed Virtual Topology based Networking (VTN) cooperated with TBCR can reduce at least 30% average end-to-end transmission delay in large-scale AANETs and provide more than 90% lower synchronization overhead than the existing solutions.

The volatile, covert and slow multistage attack patterns of Advanced Persistent Threat (APT) present a tricky challenge of APT detection, which are vital for organisations to protect their critical assets. In this article, we aim to develop system that aggregates and uses existing systems’ alerts to detect APTs. In order to achieve this, we propose a causal correlation aided semantic analysis system, calledPoirot, for detecting the multi-stage threats over a long-time span from existing systems’ alerts.Poirotis capable of autonomously mining causality between anomalous events, which instructs us in reorganizing the original alerts and in constructing alert-chains. The system further exploits the Latent Dirichlet Allocation (LDA) to model the semantic context of the alert-chains. This LDA model facilitates us to carry out the semantic analysis for capturing the latent attack intent as well as for reconstructing the APT scenario. We use an alert dataset provided by a cyber security company to verify the proposedPoirotin terms of the detection accuracy and the capability of attack scenario reconstruction. The experiment results are presented to show the achievable performance of the proposed semantic analysis based APT detection.

Low Earth Orbit (LEO) satellite networking has been an indispensable and promising concept for extending the Internet coverage of future space-air-ground integrated networks to oceanic and remote airspace. However, the topology dynamics of the LEO Satellite Network (LEO-SN) for network state perception (Liet al., 2019) and the intermittent nature of the Inter-Satellite Links (ISLs) for multipath routing discovery (Wanget al., 2019, Jianget al., 2019) both induce new complicated challenges to multipath traffic scheduling for the sake of improving the utility of the LEO-SN facility (Songet al., 2014, Zhanget al., 2018, and Yanget al., 2020). Motivated by these challenges, this paper aims to develop an AI aided intelligent multipath traffic scheduling approach for bolstering autonomous and efficient communications of LEO-SN. To achieve this, we formulate the multipath traffic scheduling problem into a pheromone incentivized Markov Decision Process (MDP) by considering ant routing protocol and adapting pheromone to LEO-SN. Employing enhanced pheromone characterizing network state, we propose ant-inspired multipath routing discovery, which is capable of promptly discovering routing paths available in the dynamic topology. To improve the utility of these discovered routing paths, we employ deep deterministic policy gradient into the pheromone-incentivized MDP-based scheduling problem to derive an intelligent multipath traffic scheduling strategy. The experimental results are further presented to show the achievable performance improvement.

The soaring mobile data traffic demands have spawned the innovative concept of mobile edge caching in ultra-dense next-generation networks, which mitigates their heavy traffic burden. We conceive cooperative content sharing between base stations (BSs) for improving the exploitation of the limited storage of a single edge cache. We formulate the cooperative caching problem as a partially observable Markov decision process (POMDP) based multi-agent decision problem, which jointly optimizes the costs of fetching contents from the local BS, from the nearby BSs and from the remote servers. To solve this problem, we devise a multi-agent actor-critic framework, where a communication module is introduced to extract and share the variability of the actions and observations of all BSs. To beneficially exploit the spatio-temporal differences of the content popularity, we harness a variational recurrent neural network (VRNN) for estimating the time-variant popularity distribution in each BS. Based on multi-agent deep reinforcement learning, we conceive a cooperative edge caching algorithm where the BSs operate cooperatively, since the distributed decision making of each agent depends on both the local and the global states. Our experiments conducted within a large scale cellular network having numerous BSs reveal that the proposed algorithm relying on the collaboration of BSs substantially improves the benefits of edge caches.

This paper considers the anti-jamming optimization problem for tracking multiple moving target flight vehicles in the presence of deception jammers. Since the radar is not able to separate the real target vehicles from a large number of deceptive vehicles, we promote the existing non-anti-jamming tracking model to the anti-jamming partially observable Markov decision process-based (POMDP-based) game tracking model by establishing a new anti-jamming Bayesian tracker. The proposed tracker is able to separate the hidden real target vehicles and establish their accurate trajectories, but the limited radar resources will decrease the accuracy. In order to effectively utilize the limited resources to guarantee the anti-jamming performance, this work deduces the anti-jamming performance gradients with respect to the resource management policy, which can be estimated with the asymptotically vanished biases. With the gradient estimates, the optimal anti-jamming resource management policy can be found with the tolerable complexity. The convergence analysis shows that the algorithm converges to a Nash equilibrium solution with probability 1. Numerical results show that the proposed algorithm can obtain the accurate target trajectories in the presence of jammers.

Promptly discovering unknown network attacks is critical for reducing the risk of major loss imposed on organizations and information infrastructure. This paper aims at developing an intelligent intrusion detection system capable of classifying known attacks as well as inferring unknown ones. To achieve this, we formulate the problem of fine-grained known/unknown intrusion detection as a two-stage minimization problem, where the first stage is to seek a score measure for minimizing the empirical risk of misclassifying the known attacks, while the second stage is to find another score measure for minimizing the identification risk of inferring unknown attacks. The hierarchical nature of problem formulation allows us to employ the class conditioned auto-encoders to construct a hierarchical intrusion detection framework. Since the reconstruction errors of unknown attacks are generally higher than that of the known attacks, we further employ extreme value theory in the second stage to model the distribution of reconstruction errors for differentiating known/unknown attack. To further reduce the false positive rate, we add a benign clustering module for learning the multimodal distribution of benign traffic. We conduct an experiment on two widely used datasets for assessing intrusion detection. The results show that the proposed method improves the detection rate of unknown attacks while keeping a low false positive rate.

Both wireless small-cell communications and software-defined networking (SDN) in wired systems continue to evolve rapidly, aiming for improving the quality of experience (QoE) of users. Against this emerging landscape, we conceive scalable video streaming over SDN-aided dense smell-cell networks by jointly optimizing the video layer selection, the wireless resource allocation, and the dynamic routing of video streams. In the light of this ambitious objective, we conceive a dense software-defined small-cell network architecture for the fine-grained manipulation of the video streams relying on the cooperation of small-cell base stations. Based on this framework, we formulate the scalable video streaming problem as maximizing the time-averaged QoE subject to a specific time-averaged rate constraint as well as to a resource constraint. By employing the classic Lyapunov optimization method, the problem is further decomposed into the twin sub-problems of video layer selection and wireless resource allocation. Via solving these sub-problems, we derive a video layer selection strategy and a wireless resource allocation algorithm. Furthermore, we propose a beneficial routing policy for scalable video streams with the aid of the so-called segment routing technique in the context of SDN, which additionally exploits the collaboration of small-cell base stations. Our results demonstrate compelling performance improvements compared with the classic PID control theory-based method.

Scalable Video Coding (SVC) has been raised as a promising technique to enable flexible video transmission for mobile users with heterogeneous terminals and varying channel capacities. In this paper, we design an adaptive layer switching algorithm for on-demand scalable video service based on receiver's buffer underflow probability (BUP). Since the low quality of channel may lead to a low buffer fullness, the buffer fullness is an indicator for reflecting the channel condition and we define BUP for characterizing the mismatch between the video bitrate and the channel throughput. Accordingly, the adaptive SVC transmission problem is formulated as the adaptive adjustment of video layers based on BUP. This allows us to optimize the attainable video quality, while keeping BUP below a desired level. To estimate BUP, we derive an analytical model based on the large deviation principles. Then, an online layer switching algorithm is proposed using this estimation model, which is capable of accommodating different channel qualities without any prior knowledge of the channel variations and of the video characteristics. We further introduce a perturbation-based layer switching approach for reducing the quality fluctuating issue caused by frequent layer switches, thus improving the viewer's quality of experience. A system prototype is implemented to evaluate the success of the proposed method. We also conduct simulations in multiuser scenarios with real video traces and the results demonstrate that the proposed algorithm is capable of improving the playback experience, while keeping a low playback interruption rate and quality variation.

In video streaming service, multicast mode is a promising way to complement unicast delivery of content, since it deliveries the video content to a broad range of receivers. It is considered as an efficient scheme to reduce redundant traffic in the networks. In this paper, we propose an OpenFlow based architecture for implementing Scalable Video Coding (SVC) multicast streaming. It enables in-network identifying, processing and manipulating the media streams, which makes prompt bitrate adaptation possible in response to network fluctuations. The heterogeneous video quality demand from the heterogeneous device also can be satisfied by customizing the multicast traffic through a centralized OpenFlow controller. We implement a testbed following the proposed architecture in our campus. With OpenFlow, we deploy IP multicast in a new way without Internet Group Management Protocol (IGMP) or any multicast addresses. Experiments implemented on the testbed show that our approach can provide a flexible and controllable video multicast streaming service and improve the usage of bandwidth resource in the condition of guaranteeing multicast receivers' Quality of Experience (QoE).