Large language models (LLMs) are increasingly used for hardware and firmware code generation, but existing studies primarily evaluate functional correctness while largely overlooking security. However, LLM-generated code that appears functionally sound may embed security flaws which could induce catastrophic damages after deployment. This critical research gap motivates us to design a benchmark for assessing security awareness under realistic specifications. In this work, we introduce HardSecBench, a benchmark with 924 tasks spanning Verilog Register Transfer Level (RTL) and firmware-level C, covering 76 hardware-relevant Common Weakness Enumeration (CWE) entries. Each task includes a structured specification, a secure reference implementation, and executable tests. To automate artifact synthesis, we propose a multi-agent pipeline that decouples synthesis from verification and grounds evaluation in execution evidence, enabling reliable evaluation. We evaluate diverse LLMs and find that they often satisfy functional requirements while leaving security risks. We also find that security results vary with prompting. These findings highlight pressing challenges and offer actionable insights for future advancements in LLM-assisted hardware design. Our data and code are available at https://github.com/chenqirui2002/HardSecBench.

Deep learning-based speech enhancement has advanced significantly, but it faces a critical trade-off: higher performance often comes at the cost of increased computational complexity, while reducing complexity tends to degrade enhancement quality. A specific manifestation of this is that lightweight models struggle to balance nonlinear distortion and noise suppression, severely hindering their deployment on resource-constrained edge devices. To address this issue, we propose a two-stage joint optimization method that integrates loss function tuning and post-processing, adhering to the "Preservation first, Removal later" (PRem) principle, termed PRSE. Specifically, we first incorporate the NDA (Noise, Distortion, and Artifacts) loss function into the lightweight model to moderately retain noise and mitigate speech distortion, while rendering residual noise stationary. Building on this foundation, we propose a low-complexity post-processing method that leverages classical noise estimation techniques to correct noise underestimation while preserving the model's ability to track non-stationary noise. When applied to two distinct architectures, DCTCLD and LiSenNet, both models yield significant improvements: the MOS increases by 0.44 and 0.64, respectively, while the additional computational overhead remains below 1%. These results narrow the performance gap with SOTA models and validate the method's generalizability, strengthening the practical feasibility of lightweight models in real-world scenarios.

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.

Satellite computing networks (SCNs) represent an emerging paradigm that integrates computational and storage capabilities into satellites, facilitating in-orbit data processing. This alleviates the downlink bottleneck over satellite–terrestrial links and enables the provision of ubiquitous, low-latency computing services. However, the management of the SCNs, given their constrained resources, high dynamics, and heterogeneous computing demands, ranging from latency-sensitive requests to compute-intensive task, remains a critical challenge. To overcome these challenges, we propose a novel dual-scale digital twin (DDT-SCN) system architecture and management paradigm to implement a collaborative task processing framework. This framework deploys a full-scale digital twin (FDT) in a ground data center for global, strategic optimization and deep reinforcement learning (DRL) model training, and a light-scale digital twin (LDT) distributed across satellites for real-time, low-latency decision-making. We introduce an incremental synchronization mechanism and a lightweight predictive model to enable the LDT to make robust, proactive decisions despite limited or delayed state information, with the FDT continually refining and transferring its global macropolicy as a compressed model to the LDT to guide local execution. Through comprehensive simulation analysis of a low Earth orbit satellite constellation, we demonstrate that DDT-SCN significantly reduces average task latency and concurrently improves the task success rate compared to conventional scheduling schemes, thereby validating its efficacy. Specifically, simulation results demonstrate that under high traffic loads (300 tasks/min), DDT-SCN maintains a task success rate of 97.7%, while concurrently reducing the average task latency by approximately 35% and communication overhead by 72.2% compared to conventional online DRL approaches.

Large language models (LLMs) are being increasingly integrated into practical hardware and firmware development pipelines for code generation. Existing studies have primarily focused on evaluating the functional correctness of LLM-generated code, yet paid limited attention to its security issues. However, LLM-generated code that appears functionally sound may embed security flaws which could induce catastrophic damages after deployment. This critical research gap motivates us to design a benchmark for assessing security awareness under realistic specifications. In this work, we introduce HardSecBench, a benchmark with 924 tasks spanning Verilog Register Transfer Level (RTL) and firmware-level C, covering 76 hardware-relevant Common Weakness Enumeration (CWE) entries. Each task includes a structured specification, a secure reference implementation, and executable tests. To automate artifact synthesis, we propose a multi-agent pipeline that decouples synthesis from verification and grounds evaluation in execution evidence, enabling reliable evaluation. Using HardSecBench, we evaluate a range of LLMs on hardware and firmware code generation and find that models often satisfy functional requirements while still leaving security risks. We also find that security results vary with prompting. These findings highlight pressing challenges and offer actionable insights for future advancements in LLM-assisted hardware design. Our data and code will be released soon.

Deploying the Unmanned Aerial Vehicle (UAV) formation as aerial base stations to construct aerial communication networks in hotspot areas holds considerable promise. However, planning the trajectories of UAV formation in complex and unknown environments, while ensuring obstacle avoidance and formation maintenance, presents unpreceding challenges. In this paper, we propose a hierarchical reinforcement learning-based trajectory planning algorithm for UAV formation. This algorithm implements two-timescale trajectory planning within a leader-follower control framework, where the leader UAV (LUAV) plans the shortest safe trajectory to the hotspot area on a large timescale and the follower UAVs (FUAVs) are responsible for obstacle avoidance and formation maintenance on a small timescale. The LUAV and FUAVs collaborate across different timescales to achieve joint trajectory optimization. To tackle the sparse reward problem in existing learning-based trajectory planning algorithms, we introduce an intrinsic curiosity-driven module that integrates historical information to enhance the exploration of the UAV formation in unknown environments. Our algorithm enhances the UAV formation’s capability to handle complex obstacles and maintain the formation, thereby improving overall performance. Simulation results demonstrate that our algorithm achieves complete obstacle avoidance with a 100% UAV survival rate. Compared to existing algorithms, our proposed algorithm reduces the trajectory length by 14% and improves the formation maintaining performance by over 90%.

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.

Disaggregated LLM inference service (DLIS), which decouples the compute-intensive prefill phase and the memory-intensive decode phase, enables more flexible and efficient resource usage. Existing solutions for deploying DLIS are typically designed for homogeneous environments. However, real-world production environments are becoming increasingly heterogeneous due to GPU shortages and rapid hardware evolution. Deploying DLIS in heterogeneous environments introduces three key challenges: 1) complicated GPU resource allocation caused by significant performance differences between GPUs, 2) communication bottlenecks caused by additional key-value (KV) cache transfers, and 3) dynamic inference loads caused by time-varying arrival rates and diverse task types. Existing methods adopted uniform model partitioning on heterogeneous GPUs and single-instance partitioning, suffering from heavy straggler effect and cross-node communication overheads. Additionally, they rely on reloading model parameters for dynamic tasks, which leads to significant service interruptions. To address these issues, we propose DisHelis, a high-throughput and low-latency DLIS system for heterogeneous environments. We formulate the DLIS deployment over heterogeneous environments as a hierarchical max-flow problem. This formulation jointly incorporates non-uniform model partitioning and hybrid instance partitioning to maximize DLIS throughput. Furthermore, we design a light-weight instance-switching approach to handle dynamic tasks without service interruptions. We solve it via a hierarchical alternating optimization algorithm that iteratively converges to a high-quality deployment plan. Experimental results show that DisHelis improves throughput by up to 1.63x and reduces latency by up to 2.4x over existing approaches.

The growing scale of deep neural network (DNN) has made distributed training essential, but concurrent jobs in shared clusters face serious network contention that limits overall performance. Existing centralized or host-side flow control approaches either face scalability bottlenecks or lack global awareness. To address these limitations, we propose JAFC, a job-aware flow control mechanism designed for distributed DNN training. JAFC leverages switches to support global state sharing across the cluster. Each work node periodically broadcasts its remaining communication volume and senses the progress of other nodes in a distributed manner. Based on this global view, each node adaptively adjusts its sending rate according to its relative remaining volume. This coordination helps stagger communication phases across jobs, thereby reducing communication overlap and alleviating network congestion. By simultaneously adjusting the transmission rates of individual work nodes within a job and coordinating bandwidth allocation across multiple jobs, JAFC reduces the overall completion time of the training process. Simulation results indicate that compared to the state-of-the-art flow control mechanisms, the proposed mechanism reduces the communication time by 1.53

The existing encrypted traffic classification schemes have not taken full advantage of the principles of network traffic encryption technology, resulting in poor classification accuracy and interpretability. In fact, the transmission process of encrypted flows can be divided into a handshake phase and a secure data transmission phase, each with distinct characteristics and inter-packet correlations. Therefore, an effective approach for encrypted traffic classification should employ different processing methods for each phase and then combine the results. In this paper, we propose the Two-Phase Feature Fusion (2P2F) approach for classifying encrypted network flows. In 2P2F, during the handshake phase, a convolutional neural network (CNN) processes the encryption protocol extension field to extract application-related features. For the secure data transmission phase, a transformer model captures content-related features from the TCP segment length sequence. Finally, an attention mechanism is used to fuse the refined features from both phases, allowing the classification model to focus more on important features and achieve accurate classification of encrypted network flows. The experimental results demonstrate the superiority of the proposed scheme.

As Bird’s Eye View (BEV) perception becomes crucial for autonomous driving, unmanned aerial vehicles (UAVs) leveraging their native top-down perspective and agile mobility are increasingly deployed to overcome the inherent occlusion and limited sensing range of traditional vehicle-mounted systems. However, enabling UAV-assisted BEV perception faces a computation-transmission dilemma: onboard BEV inference is impractical due to tight resource constraints, while offloading high-definition visual data to edge servers is bottlenecked by limited wireless bandwidth. In this paper, we present AUBP, an adaptive video offloading scheme for UAV-based BEV perception. First, AUBP introduces a spatio-semantic synergistic mechanism that combines edge-side spatial feedback from BEV inference with lightweight UAV-side semantic extraction to preserve BEV-critical information while reducing spatial redundancy. To dynamically coordinate this mechanism under time-varying wireless channels, we formulate the adaptive configuration process as a Markov Decision Process (MDP). Subsequently, a Dueling Deep Q-Network (Dueling DQN)-based decision-maker is constructed to solve this MDP, dynamically selecting optimal encoding parameters to achieve a balanced trade-off between transmission efficiency and perception quality. Experiments on the UAV dataset for BEV perception under dynamically fluctuating network conditions demonstrate that AUBP consistently surpasses representative baselines.

In response to the growing challenges posed by excessive redundant traffic in cache-enabled networks (cache-enabled networks), this paper presents a novel probabilistic caching strategy termed PCCNA (Probabilistic Caching based on Content Consistency and Node Attributes). A key innovation of PCCNA is its incorporation of strong content consistency between caching proxies and backend servers into the caching decision-making process, addressing a critical gap in existing caching strategies. Unlike traditional approaches that often focus on a singular optimization criterion, PCCNA integrates multiple factors such as content popularity within the network and the centrality of caching proxies based on their attributes. By adaptively computing the caching probability for content of varying popularity while ensuring strong consistency, PCCNA enhances the efficiency of in-network caching. Extensive simulation results demonstrate that PCCNA outperforms established caching strategies in terms of cache hit ratio, server load, and user service latency across diverse network topologies. This research provides significant insights into the development of effective caching mechanisms, paving the way for optimized resource utilization and improved user experiences in cache-enabled networks environments.

Multicast is an effective technology for improving the quality of experience (QoE) for multiple users in 360 degrees video services. However, unlike traditional video, 360 degrees video multicast presents several unique challenges, including efficient user grouping, joint optimization of resource allocation and bitrate decision, and the low transmission delay and high computing demands in the transmission. To address these issues, this article proposes a novel two-tier grouping 360 degrees video streaming multicast scheme. We first formulate an optimization model that ensures QoE while jointly optimizing multiuser two-tier grouping, resource allocation, and bitrate decisions. Subsequently, we introduce a two-tier grouping strategy and a corresponding resource allocation approach for each tier, leveraging the Shapley value for fair resource distribution. We then propose a heuristic algorithm, named two-tier grouping multicast (TTGM), which can achieve an optimal solution in certain iterations with low time complexity. Simulation results demonstrate the effectiveness of TTGM, showing that it significantly outperforms the existing algorithms, such as BF, oneG, RTOP, VG, and Dragonfly, in achieving optimal QoE performance regardless of the setting of scenarios.

Mobile edge computing (MEC) is pivotal for delivering low-latency services to IoT applications. However, inherent uncertainties in real-world applications, notably conditional execution workflows and variable service durations, are often overlooked, hindering efficient resource utilization and optimization flexibility. This paper introduces UTOC, an uncertainty-aware framework that collaboratively optimizes the execution of conditional directed acyclic graph (CDAG) applications in MEC networks by co-designing service deployment and task scheduling. First, we model service duration variability with the G/G/c queueing model and represent conditional workflows with probabilistic branching. Then we employ moment matching theory to approximate application end-to-end (E2E) latency, enabling the accurate quantification of the impact on E2E latency by service instances. This quantification enables the CDAG application execution optimisation problem to be decomposed into two subproblems, service deployment and task scheduling. Finally, we design a heuristic service deployment algorithm for efficient resource allocation and an online scheduling algorithm to minimize E2E latency while balancing server congestion within the MEC networks. Extensive experiments validate that UTOC effectively utilizes edge resources and significantly reduces E2E latency compared to state-of-the-art baselines.

In shared autonomous vehicle systems (S-AVS), efficient and adaptive vehicle repositioning plays a crucial role in meeting time-varying traffic demand, typically achieved by leveraging user demand prediction. However, most existing studies treat traffic demand prediction and shared autonomous vehicle (SAV) scheduling as separate tasks, ignoring the tight interaction between the two components, such as the potential impact of scheduling results on demand prediction. Such a design lacks a deep integration for both, potentially leading to inaccurate predictions and less efficient repositioning performance. To address this issue, we propose DRiVe, an enhanced coupling design for Demand prediction and Repositioning for shared autonomous Vehicle system. Two corresponding coupling strategies are designed, differentiated by their respective coupling locations. Specifically, we consider electric SAVs and employ model predictive control (MPC) to develop the repositioning strategy, aiming to minimize the operator's repositioning costs and passenger dissatisfaction. An online traffic demand prediction mechanism is introduced to transform scheduling actions into additional traffic demand. This additional demand is then incorporated into traditional traffic demand prediction to enhance the accuracy of the final demand prediction. The numerical results indicate that the proposed DRiVe method outperforms existing approaches in reducing passenger waiting times and repositioning distances.

Multipath congestion control for data center networks has been a research hotspot, but fully utilizing the bandwidth resources of the multipath set remains a significant challenge. This paper proposes a multipath congestion control algorithm based on in-network telemetry (INT), called INTMCC, which leverages detailed link load information to achieve joint optimization of congestion control and load balancing strategies. In INTMCC, in-network telemetry enables the acquisition of both the residual bandwidth of idle paths and the congestion level of congested paths. To achieve more accurate and smoother adjustment of the congestion control window, we design a window adjustment strategy based on the average load of idle paths and that of congested paths. Considering the varying residual bandwidth across idle paths, we derive a fine-grained load balancing scheme that aims to minimize data transmission time within a short time interval. This scheme enables us to obtain the probability of using each entropy and effectively utilize the heterogeneous residual bandwidth across different paths. Given the varying congestion levels across different paths, we assign customized suspension duration to each path to prevent further congestion exacerbation or link underutilization. Simulation results in the htsim environment show that INTMCC achieves significant performance improvements compared to STrack and SMaRTT-REPS in scenarios such as link failures and link degradation, reducing flow completion time by up to 9

The growing scale of large language models (LLMs) has made communication overhead a critical bottleneck in distributed training, primarily due to imbalanced traffic loads. Existing load balancing methods often lead to severe flow contention when handling low-entropy and high-volume LLM training flows. Motivated by the point-to-point pattern in each collective communication phase and the inherent periodicity of training traffic, we propose PhOrch, a proactive phase-level contention-free flow path orchestration framework tailored for LLM training workloads. We formulate the orchestration as an optimization problem, which is typically NP-hard. To tackle this problem, we develop a segmented edge coloring algorithm for bipartite multigraphs, which efficiently assigns flow paths while avoiding contention. Evaluation results demonstrate that PhOrch reduces the per-cycle communication time by 60% compared to the state-of-the-art methods and achieves contention-free training traffic in non-oversubscribed topologies, indicating that our method substantially mitigates the communication bottleneck during the LLM training process.

Super resolution (SR) has been proposed to reduce the bandwidth overhead and improve the user’s quality of experience (QoE) for panoramic video. However, video reconstruction greatly increases the energy consumption on mobile devices with limited battery capacity, which is rarely considered in the existing works. In this work, we propose FAESR, a Fine-grained bitrate Adaptation method with an Energy-aware Super-Resolution to maximize the QoE and minimize the energy consumption. We propose an SR power model, which is the first model to evaluate the power consumption of SR on mobile devices through actual measurements. We formulate a joint optimization problem for QoE- and energy-aware panoramic video streaming. Most neural-enhanced panoramic streaming methods use coarse-grained adaptation, either selecting only download bitrate or assigning uniform bitrate to within the predicted field of view (FoV). This can lead to bandwidth waste due to overrated tiles incorrectly predicted to be within the FoV. We develop a fine-grained bitrate adaptation algorithm based on branching sequential DRL, which jointly optimizes download and reconstruction bitrates at the tile level. Evaluation results demonstrated that FAESR can significantly reduce energy consumption by 28% while improving the QoE by 12% compared to the existing state-of-the-art works.

Virtual reality (VR) technology is widely employed across various domains, with its applications expanding as Artificial Intelligence (AI) technologies advance, bringing new scenarios and functionalities to 360-degree video streaming. However, these applications impose substantial demands on computational power and communication capacity, along with increased sensitivity to computation and transmission delay. In this paper, we propose a joint optimization scheme of computing and multicasting for 360-degree video in Mobile Edge Computing (MEC) networks to maximize the long-term Quality of Experience (QoE) of multiple users. We solve the computing and multicasting optimization problem separately by dividing it into adaptive grouping and resource optimization subproblems. By implementing adaptive grouping, we reduce redundant computation and transmission, thereby improving the efficiency of limited resource utilization. We propose a Cooperative Bargaining Game (CBG)-based resource allocation algorithm for efficient resource management and a Lyapunov optimization-based bitrate adaptation algorithm for long-term performance optimization, enhancing users' QoE while minimizing playback freezing. Our experimental results demonstrate significant improvements in multi-user long-term QoE, average bitrate, and reduced rebuffering time, underscoring the effectiveness of our scheme in demanding scenarios.