Shuochen Chang, Tong Bai, Xiaofeng Zhang, Qianli Ma, Qingyang Liu, Zhaohe Liao, Yibo Miao, Li Niu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Collaborative inference (Co-inference) across end devices and edge servers has emerged as a promising approach to satisfy the growing demand for computationally intensive and latency sensitive deep learning tasks. However, the limited computational capabilities and the bandwidth constraints of devices and edge servers pose significant challenges on co-inference. Current pertinent solutions either rely on predefined partition points and compression ratios or support only limited dimensions of dynamic adjustment, resulting in insufficient flexibility to adapt to diverse user requirements and network conditions. To address these challenges, we propose a novel framework that enhances co-inference through adaptive intermediate feature compression and efficient resource allocation. Specifically, we design a sophisticated feature compression method that incorporates channel pruning, spatial downsampling, and quantization, enabled by a weight-shared dynamic neural network architecture for efficient compression parameter switching without model reloading. Then, we formulate a constrained accuracy-maximization problem and develop a dynamic programming-based solution to jointly optimize partition points, compression parameters, and resource allocation, while meeting diverse user requirements. Experimental results show that our approach is capable of achieving up to a 22.8% improvement in inference accuracy compared to the state-of-the-art method in multi-user scenarios, demonstrating our superiority.

As a killer app of edge computing, real-time video analytics has found its wide usage in diverse applications, such as security surveillance and manufacturing automation. Unlike state-of-the-art efforts in edge video analytics, which primarily focus on single-task scenarios, we address multi-task video analytics, enabling concurrent execution of multiple tasks on a single video stream. Specifically, our approach aims to minimize monetary costs for the edge service provider through efficient query and resource scheduling, while meeting accuracy and latency requirements of diverse video analytics tasks. A crucial prerequisite for this is to determine the relationship between video analytics accuracy and system configuration parameters. We design a Transformer-aided configuration-accuracy predictor to capture both the current video content and inter-frame temporal dependencies, generating precise configuration-accuracy profiles in real-time. To better exploit the scarce communication and computing resources, a query merging technique is employed, which allows queries from the same camera to share the network bandwidth and neural network models, leading to reduced resource consumption. A heuristic algorithm is then readily proposed to schedule video queries and resources, which dynamically adapts video configurations, query merging, video analytics model selection, task placement, and GPU provisioning. Experimental results show that our system achieves near-optimal performance and outperforms state-of-the-art methods, demonstrating the superiority of our collaboration scheme.

Characterized by its ease of low-latency response, edge computing is capable of supporting real-time video analytics applications, constituting an edge video analytics paradigm, where the joint knob configuration and network scheduling design has drawn ever-escalating research attention. However, the potential of edge video analytics has not been fully exploited, owing to the limitations of the state-of-the-art as follows. i) The eminent impact of video content on accuracy performance has been ignored. ii) The variables that can be tuned are not fully considered in scheduling. iii) The heuristic algorithm-based solutions are far from the optimal. To fill in this gap, in this paper, we conceive a content-aware joint knob configuration and resource allocation scheme for edge video analytics. Concretely, fed with the features extracted from the video content, a deep neural network (DNN)-based predictor is proposed to predict the configuration-accuracy performance in a real-time manner. With an aid of the predictive results, we formulate an accuracy-maximization problem as an integer programming problem, by optimizing the variables, including resolution, frame rate, video analytic model, network bandwidth, and computational resource subject to the latency constraints. To solve this problem in an efficient manner, we devise a novel low-complexity dynamic programming method. Simulation results verify the efficiency of our content-aware joint knob configuration and resource allocation scheme. Quantitatively, a 3.3% gap is attained towards the upper bound in terms of the accuracy in an object detection scenario, relying on the scheme proposed.

With the increasing number of mobile devices, mobile crowdsensing (MCS) has garnered significant attention in research. However, computing infrastructures such as edge/cloud nodes, which are necessary for processing sensor data, are not always readily available. To address this issue, we propose a cooperative computing framework that enables the offloading of sensor data to nearby mobile devices with unused computational resources (known as helpers) for processing. Our approach considers a scenario with multiple sources and multiple helpers, where computational tasks can be partially offloaded to several helpers. We jointly optimize task offloading strategy, communication resources, and computational resources to minimize the weighted sum energy consumption of mobile devices. We model the optimization problem as a mixed- integer nonlinear programming (MINLP), with the source-helper assignment solved using a distributed algorithm based on matching theory, and the joint task partition and resource allocation problem solved using an alternating optimization (AO) method. Simulation results demonstrate the efficacy of our cooperative computing framework and scheduling scheme, which offer significant advantages over local computing in terms of reducing the weighted sum energy consumption and improving the task completion ratio.

As for the time-division communications system, the pilot spoofing attack (PSA) technique is maliciously utilized by active eavesdroppers during the uplink training phase, for contaminating the legitimate channel estimation and thus altering the beamforming design towards the eavesdroppers. Nonorthogonal multiple access (NOMA) has been recognized as the key technology for the envisioned Internet of Things (IoT) networks. In order to prevent the aforementioned information leakage in NOMA-IoT systems, we develop a novel two-way training scheme to detect PSA and a robust secure beamforming design for providing secure transmission, by utilizing the emerging technique of reconfigurable intelligent surface (RIS), which is turned off during the uplink training phase and turned on during the downlink training phase, respectively. Considering that the perfect channel state information related to the eavesdropping channel is typically difficult to obtain, a secrecy outage probability-constrained robust secure beamforming design is proposed to maximize the achievable sum secrecy rate of the legitimate users, by alternatively optimizing the active beamforming and RIS passive beamforming, while satisfying the requirements of the NOMA transmission. Elaborate simulation results reveal that the proposed detection method attains a super PSA detection performance and the proposed robust secure beamforming design is capable of efficiently enhancing the achievable sum secrecy rate, compared with various benchmark schemes.

The Internet of underwater things (IoUT) is envisioned to be an essential part of maritime activities. Given the IoUT devices’ wide-area distribution and constrained transmit power, autonomous underwater vehicles (AUVs) have been widely adopted for collecting and forwarding the data sensed by IoUT devices to the surface-stations. In order to accommodate the diverse requirements of IoUT applications, it is imperative to conceive a multi-tier underwater computing (MTUC) framework by carefully harnessing both the computing and the communications as well as the storage resources of both the surface-station and of the AUVs as well as of the IoUT devices. Furthermore, to meet the stringent energy constraints of the IoUT devices and to reduce the operating cost of the MTUC framework, a joint environment-aware AUV trajectory design and resource management problem is formulated, which is a high-dimensional NP-hard problem. To tackle this challenge, we first transform the problem into a Markov decision process (MDP) and solve it with the aid of the asynchronous advantage actor-critic (A3C) algorithm. Our simulation results demonstrate the superiority of our scheme.

Fluids injected during hydraulic fracturing (fracking) in unconventional shale oil and gas reservoirs, geothermal system enhancement, wastewater disposal, and carbon capture and storage can induce microearthquakes. The spatiotemporal distribution of induced earthquakes is often used to trace the growth of fractures in target layers and guides production. We analyze microseismicity behavior induced by fracking in the Montney Formation, one of the largest unconventional oil and gas reservoirs in North America. An optical fiber deployed in a horizontal well provides extensive spatial sampling and data coverage for microseismic imaging. We design median and F-k filters to predict instrumental and random noise and further suppress them by adaptive noise subtraction. An elliptical vertical transverse isotropic (VTI) velocity model is derived from the Backus-averaged well-log sonic data and is modified to match the microseismic wavefronts by a grid search. We image 41 previously cataloged microearthquakes recorded by DAS using Geometric-mean Reverse Time Migration. We find that the fiber geometry’s lack of 2D/3D variations increases the non-uniqueness of the image point location, and the P-wave particle motions derived from 3C geophones data can effectively eliminate the location ambiguity. The spatiotemporal distribution of our updated locations agrees with the fracking schedule. Predicted P- and S-wave traveltimes from the updated locations also match with the observed waveforms. Analyzing data sensitivity to source locations confirms the potential limitations imposed on source imaging by the geometry of borehole observations and shows that relocation accuracy is directionally dependent. We also investigate the feasibility of estimating source focal mechanisms using realistic DAS and geophone observations. Our study provides guidance for characterizing microearthquake sources and optimizing observation geometry for unconventional reservoirs.

Reconfigurable intelligent surface (RIS) is capable of improving the wireless system performance by steering the reflected signal in the desired direction. One of the major challenges is that both the transceiver and RIS have to be jointly optimized, where the optimization problems have to be reformulated for different system models and scenarios. To circumvent this challenge, new low-complexity antenna selection (AS) algorithms for transceiver-agnostic RIS configuration are proposed. Given a multiple-input multiple-output (MIMO) channel, the proposed RIS-AS opts for accurately aligning the RIS both with the transmit antenna (TA) and receive antenna (RA) for the sake of maximizing the MIMO channel’s overall output power. The proposed RIS-AS only has to configure the RIS alone, i.e. without iterations with the transceiver optimization. As a result, the proposed RIS-AS has the compelling benefit that they are generically applicable, regardless of the specific transceiver architecture. Our simulation results confirm that the proposed RIS-AS is capable of supporting any MIMO configuration, regardless of their closed/open-loop, single-/ full-RF and multiplexing-/diversity-oriented setups.

Unmanned aerial vehicle (UAV)-enabled backscatter communications (BackComm) is deemed to be a vital technique enabling the data transmission over massive battery-less devices for Internet of Things (IoT). However, the UAV-enabled Backcomm suffers from information leakage due to the broadcasting nature of wireless channels. To cope with the security issue, in this paper, a UAV-enabled multi-user secure BackComm system is developed using analog beamforming (ABF) and randomized continuous wave (RCW) techniques, where the multiple users are supported by the multi-carrier RCW over a single low-complexity radio frequency (RF) chain. By exploiting the RCW transmitted towards backscatter, the eavesdropping link can be eroded without any specific jamming signals. The closed-form of the secrecy rate is studied with the approximations, which is then maximized by jointly optimizing the beamforming together with the UAV’s location and the RCW settings. Simulation results are carried out to confirm the accuracy of the proposed approximation, while the convergence behavior of the optimization algorithm is analyzed. As a result, it can be shown that the secrecy rate can be significantly improved compared with the benchmark schemes.

Owing to their intrinsic advantages of seamless coverage and of high data rate, space-air-ground communications networks (SAGCN) are recognized as one of the emerging technologies for the future wireless communications systems. However, the broadcasting nature of wireless communications inevitably imposes security issues on SAGCN. In this paper, we consider the uplink of the full-duplex unmanned aerial vehicle (UAV)-aided three-layer SAGCN, comprising of ground Internet of Remote Things (IoRT) terminals, an unmanned aerial vehicle, and a low-earth orbit (LEO) satellite, where eavesdroppers are intercepting the information transmitted. In order to ensure a secure uplink transmission, a joint UAV deployment and power allocation scheme is conceived for maximizing the secrecy rate of the SAGCN, subject to the following constraints: i) UAV’s power, ii) the UAV deployment area, and iii) the secrecy rate, which are imposed on the different layers. More explicitly, once we formulate a joint optimization problem to maximize the secrecy rate, we decouple the variables and decompose the original problem into multiple subproblems in a tractable manner. Then, we simplify the subproblems with the aid of slack variables and solve them relying on the successive convex approximation method. Following this, initialization schemes are designed to exploit the one-direction greedy method for diverse environment settings, for speeding up the convergence of the iterative algorithm proposed. Finally, simulation results reveal that the convergence can be achieved within a small number of iterations by the proposed initialization scheme, while the algorithm conceived is capable of attaining a substantial improvement of the secrecy rate for the SAGCN.

Queuing delay is of essential importance in the Internet-of-Things scenarios where the buffer sizes of devices are limited. The existing cross-layer research contributions aiming at minimizing the queuing delay usually rely on either transmit power control or dynamic spectrum allocation. Bearing in mind that the transmission throughput is dependent on the distance between the transmitter and the receiver, in this context we exploit the agility of the unmanned-aerial-vehicle (UAV)-mounted base stations (BSs) for proactively adjusting the aerial BS (ABS)’s placement in accordance with wireless teletraffic dynamics. Specifically, we formulate a minimum-delay ABS placement problem for UAV-enabled networks, subject to realistic constraints on the ABS’s battery life and velocity. Its solutions are technically realized under three different assumptions in regard to the wireless teletraffic dynamics. The backward induction technique is invoked for both the scenario where the full knowledge of the wireless teletraffic dynamics is available, and for the case where only their statistical knowledge is available. In contrast, a reinforcement learning aided approach is invoked for the case when neither the exact number of arriving packets nor that of their statistical knowledge is available. The numerical results demonstrate that our proposed algorithms are capable of improving the system’s performance compared to the benchmark schemes in terms of both the average delay and of the buffer overflow probability.

Wireless powered mobile edge computing (WP-MEC) has been recognized as a promising technique to provide both enhanced computational capability and sustainable energy supply to massive low-power wireless devices. However, its energy consumption becomes substantial, when the transmission link used for wireless energy transfer (WET) and for computation offloading is hostile. To mitigate this hindrance, we propose to employ the emerging technique of intelligent reflecting surface (IRS) in WP-MEC systems, which is capable of providing an additional link both for WET and for computation offloading. Specifically, we consider a multi-user scenario where both the WET and the computation offloading are based on orthogonal frequency-division multiplexing (OFDM) systems. Built on this model, an innovative framework is developed to minimize the energy consumption of the IRS-aided WP-MEC network, by optimizing the power allocation of the WET signals, the local computing frequencies of wireless devices, both the sub-band-device association and the power allocation used for computation offloading, as well as the IRS reflection coefficients. The major challenges of this optimization lie in the strong coupling between the settings of WET and of computing as well as the unit-modules constraint on IRS reflection coefficients. To tackle these issues, the technique of alternating optimization is invoked for decoupling the WET and computing designs, while two sets of locally optimal IRS reflection coefficients are provided for WET and for computation offloading separately relying on the successive convex approximation method. The numerical results demonstrate that our proposed scheme is capable of monumentally outperforming the conventional WP-MEC network without IRSs. Quantitatively, about 80% energy consumption reduction is attained over the conventional MEC system in a single cell, where 3 wireless devices are served via 16 sub-bands, with the aid of an IRS comprising of 50 elements.

Attracted by the advantages of multi-access edge computing (MEC) and non-orthogonal multiple access (NOMA), this article studies the resource allocation problem of a NOMA-MEC system in an ultra-dense network (UDN), where each user may opt for offloading tasks to the MEC server when it is computationally intensive. Our optimization goal is to minimize the system computation cost, concerning the energy consumption and task delay of users. In order to tackle the non-convexity issue of the objective function, we decouple this problem into two sub-problems: user clustering as well as jointly power and computation resource allocation. Firstly, we propose a user clustering matching (UCM) algorithm exploiting the differences in channel gains of users. Then, relying on the mean-field game (MFG) framework, we solve the resource allocation problem for intensive user deployment, using the novel deep deterministic policy gradient (DDPG) method, which is termed by a mean-field-deep deterministic policy gradient (MF-DDPG) algorithm. Finally, a jointly iterative optimization algorithm (JIOA) of UCM and MF-DDPG is proposed to minimize the computation cost of users. The simulation results demonstrate that the proposed algorithm exhibits rapid convergence, and is capable of efficiently reducing both the energy consumption and task delay of users.

Computation off-loading in mobile edge computing (MEC) systems constitutes an efficient paradigm of supporting resource-intensive applications on mobile devices. However, the benefit of MEC cannot be fully exploited, when the communications link used for off-loading computational tasks is hostile. Fortunately, the propagation-induced impairments may be mitigated by intelligent reflecting surfaces (IRS), which are capable of enhancing both the spectral- and energy-efficiency. Specifically, an IRS comprises an IRS controller and a large number of passive reflecting elements, each of which may impose a phase shift on the incident signal, thus collaboratively improving the propagation environment. In this paper, the beneficial role of IRSs is investigated in MEC systems, where single-antenna devices may opt for off-loading a fraction of their computational tasks to the edge computing node via a multi-antenna access point with the aid of an IRS. Pertinent latency-minimization problems are formulated for both single-device and multi-device scenarios, subject to practical constraints imposed on both the edge computing capability and the IRS phase shift design. To solve this problem, the block coordinate descent (BCD) technique is invoked to decouple the original problem into two subproblems, and then the computing and communications settings are alternatively optimized using low-complexity iterative algorithms. It is demonstrated that our IRS-aided MEC system is capable of significantly outperforming the conventional MEC system operating without IRSs. Quantitatively, about 20 % computational latency reduction is achieved over the conventional MEC system in a single cell of a 300 m radius and 5 active devices, relying on a 5-antenna access point.

With the dramatic growth of seismic data volume, efficient and accurate seismic source location has become a significant challenge to seismologists. Recently, time reversal imaging (TRI) has been widely applied in automatic seismic source location for its robustness and accuracy, but its wave-equation-based implementation is usually computationally expensive. To achieve an efficient in situ and real-time source location, the emerging sensor network is a good option. In this article, we propose a simplified Gaussian beam TRI (SGTRI) method to implement the seismic source location in a distributed sensor network. Gaussian beam (GB) is a high-frequency asymptotic solution of the wave equation, which can help reduce the computation costs of the wavefield extrapolation in conventional TRI. Traditionally, the GB construction for reflection seismic imaging covers the entire subsurface space. However, for certain source localization, only limited areas contribute. Thus, we propose a beamforming-technique-based simplified GB construction to further boost efficiency. Then, we propose an imaging condition for the SGTRI to construct the final source location map. Using synthetic experiments, we demonstrate the accuracy, robustness, and efficiency of the proposed method compared with conventional TRI. In the end, a field application also shows promising results.

Unmanned aerial vehicles (UAVs) have been widely used in a range of compelling applications. However, some of them are incompetent in tackling with computation-intensive tasks due to limited processing capability and battery life. In this paper, we combine the mobile edge computing and traditional cloud computing techniques for offloading the tasks from multi-UAV systems. Specifically, we jointly optimize the task scheduling and resource allocation in the heterogeneous cloud architecture, where we strike a power-delay trade-off of the system relying on the queue theory and Lyapunov optimization, followed by its optimal strategy analysis in each time slot. Moreover, we conceive an iterative algorithm with a closed-form solution at each iteration round in order to reduce the computational complexity. Finally, numerical results demonstrate both the feasibility and effectiveness of our proposed scheme. This paper validates that the heterogeneous cloud structure can be the beneficial for improving quality-of-service performance of multi-UAV systems.

In this treatise, first of all, we conceive a generic multiple-symbol differential sphere detection (MSDSD) solution for both single- and multiple-antenna-based noncoherent schemes in both uncoded and coded scenarios, where the high-mobility aeronautical Ricean fading features are taken into account. The bespoke design is the first MSDSD solution in the open literature that is applicable to the generic differential space-time modulation (DSTM) for transmission over Ricean fading. In the light of this development, the recently developed differential spatial modulation and its diversity counterpart of differential space-time block coding using index shift keying are specifically recommended for aeronautical applications owing to their low-complexity single-RF and finite-cardinality features. Moreover, we further devise a noncoherent decision-feedback differential detection and a channel-state information estimation aided coherent detection, which also take into account the same Ricean features. Finally, the advantages of the proposed techniques in different scenarios lead us to propose for the aeronautical systems to adaptively: 1) switch between coherent and non-coherent schemes; 2) switch between single- and multiple-antenna-based schemes as well as; and 3) switch between high-diversity and high-throughput DSTM schemes.

Unmanned aerial vehicles (UAVs) cooperating with satellites and base stations (BSs) constitute a space-air-ground three-tier heterogeneous network, which is beneficial in terms of both providing the seamless coverage as well as of improving the capacity for the users. However, cross-tier interference may be inevitable among these tightly embraced heterogeneous networks. In our paper, we propose a two-stage joint hovering altitude and power control solution for the resource allocation problem. Furthermore, Lagrange dual decomposition and concave-convex procedure (CCP) method are used to solve this problem. Finally, simulation results show the effectiveness of our proposed two-stage joint optimization algorithm in terms of UAV network's total throughput.

Body area networks (BANs) are systems of wearable computing devices for long‐term monitoring of personal health care. BAN is an emerging technology for the worldwide ageing population. In the BAN system, the transceiver is the most energy‐consuming part of a sensor node and radio transmission in the vicinity of the human body is highly lossy and inefficient. Therefore, the energy of the sensor node constraints the life cycle and quality of service of the network; consequently, low‐cost protocol shaves attracted wide interest. This study proposes a frame structure model of a self‐adaptive guard band protocol, which introduces a GB in each time slot according to the allowed maximum time drift of the crystal, adaptively adjusts the value of the GB based on the actual time drift, and then ensures that the node simultaneously maintains the sleeping state and synchronisation with the coordinator during beacon transmission, thus reducing the energy consumption.

Many state-of-the-art general object detection methods make use of shared full-image convolutional features (as in Faster R-CNN). This achieves a reasonable test-phase computation time while enjoys the discriminative power provided by large Convolutional Neural Network (CNN) models. Such designs excel on benchmarks1which contain natural images but which have very unnatural distributions, i.e. they have an unnaturally high-frequency of the target classes and a bias towards a “friendly” or “dominant” object scale. In this paper we present further study of the use and adaptation of the Faster R-CNN object detection method for datasets presenting natural scale distribution and unbiased real-world object frequency. In particular, we show that better alignment of the detector scale sensitivity to the extant distribution improves vehicle detection performance. We do this by modifying both the selection of Region Proposals, and through using more scale-appropriate full-image convolution features within the CNN model. By selecting better scales in the region proposal input and by combining feature maps through careful design of the convolutional neural network, we improve performance on smaller objects. We significantly increase detection AP for the KITTI dataset car class from 76.3% on our baseline Faster R-CNN detector to 83.6% in our improved detector.