With the development of IoT technologies, the demand for quality of service across different business scenarios has become increasingly diverse. Currently, how to accurately interpret IoT service requirements and effectively guarantee quality of service (QoS) for differentiated business scenarios remains a significant challenge in the industry. To address this challenge, in-depth research on intent-driven IoT resource allocation mechanisms is urgently needed. Given the dynamic and complex nature of IoT environments, along with the notable differences in service requirements, traditional demand interpretation methods are found to be inadequate for achieving accurate mapping from user intents to IoT policies. Therefore, an intent-driven dual-time-scale resource allocation method was proposed. Firstly, scenarios were categorized based on their QoS requirements, and an intent translation algorithm based on a large language model with retrieval enhancement was designed. By integrating device and service information, user intents were translated into executable IoT policies. Secondly, focusing on both resource allocation efficiency and the dynamic state changes in IoT environments, a dual-time-scale resource allocation framework was constructed with the optimization objective of maximizing the overall system revenue. This framework included: at the long time scale, the dueling double deep Q-network (D3QN) algorithm was adopted for global resource allocation among slices; at the short time scale, the mixed-integer linear programming (MILP) algorithm was employed for fine-grained resource scheduling within slices. Experimental results demonstrated that the proposed method enabled accurate generation of upper-level policies through intent translation and achieved elastic scheduling of lower-level resources via the resource allocation algorithms. Compared with single-time-scale algorithms, the proposed approach allowed for more efficient resource allocation that better adapted to the state of downstream IoT systems, thereby ensuring differentiated QoS for various business scenarios within the same IoT environment.

Federated Graph Learning enables multiple clients to collaboratively train graph models while protecting local private data. However, most studies have assumed that all clients contribute data voluntarily and actively. Without reasonable incentives, clients are often reluctant to contribute personal data for model training. Furthermore, the budget for incentives is limited, and if clients with low-quality graph data are incentivized to participate in training, it will negatively impact the training performance of all parties in the system. To address this, we propose AEFGL, a Reverse Auction and Value Evaluation-Based Incentive Mechanism for Federated Graph Learning. First, we design a reverse auction mechanism combining graph structural attribute motifs with client production value. Then, we propose a method for evaluating client production value based on the comparison of the client's expected reward and actual value. This mechanism can incentivize clients with high-quality graph data to participate in training within budget constraints, thereby improving the model quality. Experimental results validate the superiority of the AEFGL mechanism and the economic properties it satisfies.

Graph Federated Learning (GFL) has shown great potential in privacy protection and distributed intelligence through distributed collaborative training of graph-structured data without sharing raw information. However, existing GFL approaches often lack the capability for comprehensive feature extraction and adaptive optimization, particularly in non-independent and identically distributed (NON-IID) scenarios where balancing global structural understanding and local node-level detail remains a challenge. To this end, this paper proposes a novel framework called GFL-SAR (Graph Federated Collaborative Learning Framework Based on Structural Amplification and Attention Refinement), which enhances the representation learning capability of graph data through a dual-branch collaborative design. Specifically, we propose the Structural Insight Amplifier (SIA), which utilizes an improved Graph Convolutional Network (GCN) to strengthen structural awareness and improve modeling of topological patterns. In parallel, we propose the Attentive Relational Refiner (ARR), which employs an enhanced Graph Attention Network (GAT) to perform fine-grained modeling of node relationships and neighborhood features, thereby improving the expressiveness of local interactions and preserving critical contextual information. GFL-SAR effectively integrates multi-scale features from every branch via feature fusion and federated optimization, thereby addressing existing GFL limitations in structural modeling and feature representation. Experiments on standard benchmark datasets including Cora, Citeseer, Polblogs, and Cora_ML demonstrate that GFL-SAR achieves superior performance in classification accuracy, convergence speed, and robustness compared to existing methods, confirming its effectiveness and generalizability in GFL tasks.

In this paper, we investigate knowledge forgetting in large language models with a focus on its generalisation—ensuring that models forget not only specific training samples but also related implicit knowledge. To this end, we begin by identifying a broader unlearning scope that includes both target data and logically associated samples, including rephrased, subject-replaced, relation-reversed, and one-hop reasoned data. We then conduct a rigorous evaluation of 15 state-of-the-art methods across three datasets, revealing that unlearned models still recall paraphrased answers and retain target facts in their intermediate layers. This motivates us to take a preliminary step toward more generalised implicit knowledge forgetting by proposing PERMU—a novel probability perturbation-based unlearning paradigm. PERMU simulates adversarial unlearning samples to eliminate fact-related tokens from the logit distribution, collectively reducing the probabilities of all answer-associated tokens. Experiments are conducted on a diverse range of datasets, including TOFU, Harry Potter, ZsRE, WMDP, and MUSE, using models ranging from 1.3B to 13B in scale. The results demonstrate that PERMU delivers up to a 50.40% improvement in unlearning vanilla target data while maintaining a 40.73% boost in forgetting implicit knowledge. Our code can be found in the supplementary material.

Massive Internet of Things (IoT) devices facilitate real-time monitoring and precise decision-making through the unceasing transmission of status update flows. Nevertheless, when multiple new devices access and compete, the CPU and bandwidth resources allocated to existing traffic by the User Plane Function (UPF) are significantly reduced. Such a sudden reduction causes the queue length in the UPF to spike, resulting in packets not being forwarded timely, a phenomenon known as transient congestion. Consequently, tail latency experiences a substantial increase, and the update process is significantly disrupted. However, constrained by the substantial proliferation of interconnected IoT devices and the inflated congestion control loop, existing solutions are unable to adequately address transient congestion in the UPF. In this paper, we propose HyperWay, a tunnel-based solution that provides stable UPF service for IoT applications by synergizing compressing and scheduling capsules. To timely forward a maximum number of packets within the constrained resources, HyperWay establishes QUIC-enabled tunnels that support the compression of packets into capsules. Meanwhile, HyperWay schedules capsules with more urgent status updates for priority forwarding through low-complexity computing, based on the results of mathematical analysis. We have implemented and evaluated HyperWay in both a software-defined platform and a real-world 5 G network, while results show that HyperWay significantly reduces latency spikes and status update performance degradation by 12% to 89%.

Hosted retrieval-augmented generation (RAG) and semantic search allow users to query valuable provider-held corpora, raising two competing demands: to hide each query and chosen result, yet reveal only the documents that the user is authorized to receive. Existing cryptographic approaches either make this costly by processing the entire corpus for every query, or sacrifice quality for efficiency by scanning a few clusters. We repurpose learned deep hashing as a private filter: a randomized binary code points the provider to a short candidate list, while encrypted reranking and oblivious key transfer protect the precise query and final selection. This shortlist short-circuits full-corpus cryptographic search without sacrificing retrieval quality: with 200-500 candidates, it closely matches full-corpus retrieval across five zero-shot corpora spanning 25K to 5.4M documents. On the full 2.68M-passage NQ corpus over a 10-Gbps link, our protocol only adds 0.73 seconds, or 10 percent, to a 128-token Qwen3-32B RAG pipeline. The released code satisfies directional metric differential privacy (DP) and substantially reduces embedding-inversion and property-inference leakage, demonstrating that a carefully learned shortlist can make private dense retrieval both accurate and practical.

Modern networks require frequent control plane changes, which make it challenging to analyze the differences and similarities (i.e., relations) between network configurations before and after changes, as well as to identify the impacted specifications. Relational network verification effectively overcomes these problems. However, current tools still suffer from several limitations, including semi-automation, prolonged verification time, poor scalability, and limited expressiveness.To address these issues, this paper presents Relational Network Verification using Reinforcement Learning (RNV-RL). This framework integrates a specialized Domain-Specific Language (DSL) called Enhanced Regular Intermediate Relations (ERIR) with Reinforcement Learning (RL) techniques to automate relational network verification. RNV-RL translates input change intents and network configuration pairs into ERIR expressions, which compactly and comprehensively derive relational specifications from change intents, thereby enhancing enhance expressiveness. Building on this foundation, RNV-RL employs RL techniques to reduce verification time and improve scalability. It formulates the verification process as an optimization task for strategy search within a task-distribution Markov Decision Process (MDP) framework, where strategies are modeled as decision trees. A two-step approach is proposed to identify optimal strategies: the Multi-Distribution Policy Gradient (MD-PG) algorithm predicts the most promising strategies, and then the Policy-Guided Relational Network Verification Search (PG-RNVS) algorithm prioritizes and refines them.We compare RNV-RL with the latest relational network verification tool Rela and two single-snapshot verification tools, Batfish and Minesweeper. The evaluation results show that RNV-RL can express 100% of change intents while Rela only supports 97%. In contrast, Batfish and Minesweeper fail to extract change intents. Furthermore, RNV-RL is 10 times faster than Rela for 80% of the changes, while the other two tools are prone to timeouts. These performance gains are consistently observed across various specification sizes and location granularity levels.

This paper proposes a robust design for a free-space optical (FSO) system assisted by an unmanned aerial vehicle (UAV) equipped with an intelligent reflecting surface (IRS), operating under probabilistic malicious jamming. The UAV-carried IRS establishes an auxiliary link when the direct path is blocked. The system experiences composite fading (saturated turbulence, pointing errors, and angle-of-arrival fluctuations), and the jammer's intermittent activity is modeled by a Bernoulli process. We derive a closed-form expression for the average outage probability (OP) under this unified channel-and-jamming model. To address the critical performance-cost trade-off, a bi-objective optimization problem is formulated to jointly minimize the OP and the hardware deployment cost. An alternating optimization (AO) algorithm is proposed to solve the resulting mixed-integer nonlinear programming problem by decoupling it into discrete (number of IRS elements) and continuous (power, angles, apertures) subproblems, which are efficiently handled via integer search and particle swarm optimization, respectively. Simulation results demonstrate that the proposed AO algorithm converges within 15 iterations, substantially faster than genetic algorithm and random search, and achieves a near-optimal trade-off, reducing outage probability by an order of magnitude compared to non-optimized benchmarks while keeping hardware cost within budget.

Network configuration management has become increasingly complex, inefficient, and prone to errors due to frequent updates in command structures and the prevalence of multi-vendor network infrastructures. To tackle these challenges, this paper introduces a novel cognitive communication approach, formulating a new task called intent-driven multi-vendor network configuration generation. Within the broader intent-based networking lifecycle, this task specifically targets the realization and command generation stage—translating natural language operational intents into accurate and syntactically valid network commands compatible with multiple vendors, rather than addressing high-level intent interpretation or decomposition. Three primary challenges are addressed: syntactical command validity, vendor-specific syntax diversity, and outdated or inconsistent network knowledge. We propose ConfGen, a cognitive and intent-driven multi-vendor configuration generation framework that consists of two phases: vendor-agnostic syntax retrieval and syntax-constrained command generation. In the first phase, a cognitive retrieval mechanism and reranking strategy identify the most relevant syntax structures based on user intents, while vendor-specific syntax components are effectively generalized. The second phase employs a Large Language Model (LLM) guided by retrieved syntax constraints and user intents to generate precise and valid network commands. To ensure syntactical correctness and vendor compatibility, syntax-constrained decoding strategies are integrated into the LLM generation process. Extensive experimental evaluations conducted on a novel dataset containing network commands from Huawei, Cisco, Nokia, and Juniper demonstrate the superiority of ConfGen. Results confirm significant performance improvements over state-of-the-art solutions in generating accurate, multi-vendor-compatible network configurations driven by user intent.

Training large visual neural networks is highly memory-intensive, with optimizer states often acting as the main bottleneck. While memory-efficient subspace optimizers (MeSO) alleviate this by projecting gradients into low-dimensional subspaces, they severely limit updates to restricted subspaces—this degrades optimization expressivity and traps pretraining in suboptimal regions, a critical issue for visual tasks requiring broad parameter exploration. To address this long-standing limitation, we propose ViLaX(Visual Low-rank to High-rank Update eXtension), a novel dual-factor reparameterization framework that achieves low-rank gradient compression while decoupling the optimizer update rank. By factorizing each weight matrix into two low-rank components and applying independent MeSO updates to each, ViLaX enables high-rank effective updates to the original weights, a key innovation that preserves memory efficiency without sacrificing optimization capability. Extensive experiments on CIFAR-10/100 and Tiny ImageNet demonstrate ViLaX’s superiority: it outperforms state-of-the-art low-rank MeSO methods in pretraining accuracy, solving the critical memory-expressivity tradeoff that has challenged visual neural network training.

Blockchain has gained increasing attention as a promising technology for reliable communication and secure collaboration in various fields. However, deploying blockchain in communication systems remains challenging due to its limited performance and large costs. To address this challenge, we propose blockchain generative intelligence (BGI) to enhance both scalability and security of blockchain-based communication systems. Firstly, we introduce the three-level optimization for blockchain enhancement. Based on this, we introduce how to leverage Generative AI (GenAI) to address the scalability and security issues within the three levels of blockchain optimization. Furthermore, in the BGI framework, we incorporate existing blockchain scaling strategies and utilize the advantages of GenAI to mitigate the challenges faced by these scaling strategies, including directed acyclic graph (DAG), sharding, protocol optimization and cross-chain technology. Moreover, we conduct a case study demonstrating the effectiveness of deploying BGI-assisted sharding blockchain in autonomous vehicle platooning (AVP), as one of the promising application scenarios of blockchain. Finally, we evaluate the performance of the proposed strategies by simulations.

This work proposes two approaches, i.e., Fed-knowledge distillation (KD) and Fed-tensor knowledge distillation (TKD), to enhance communication efficiency with federated learning (FL) in Industrial Internet of Things (IIoT) radar sensing systems, which face the challenges in transmitting large volumes of sensitive data while still ensuring privacy. Fed-KD leverages KD to transfer knowledge from complex teacher networks to simpler student models in order to reduce communication overhead in bandwidth-constrained environments. Fed-TKD improves communication efficiency further by applying tensor decomposition to reduce parameter redundancy. Experimental results using an IIoT radar imagery dataset show that both methods can significantly reduce communication costs while maintaining a high model accuracy, making them especially suitable for privacy-preserving IIoT sensing applications. Furthermore, experimental results with both independent and identically distributed (IID) and non-IID data distributions confirm the robustness of the proposed methods in heterogeneous environments.

Network configuration synthesis is essential for automated configuration management in large and complex networks. However, existing synthesizers face challenges in practical applications, including limited scalability, slow synthesis speed, insufficient support for various routing protocols, and difficulty in handling mixed vendor configurations. To address these issues, this paper proposes MoCS, a modular configuration synthesizer that integrates multiple Large Language Models (LLMs) with Graph Neural Network (GNN)-enhanced recommendations to enable protocol-agnostic and vendor-compliant configuration synthesis. MoCS decomposes the synthesis pipeline into three LLM-based modules, each following a unified prompt engineering framework with task-specific adaptations. Specifically, the Intent Translation Module (IT-Module) translates natural language intents into structured configuration tasks, while the Configuration Graph Generation Module (CG-Module) constructs a Configuration Knowledge Graph (CKG) by incorporating semantic information from network topologies, structured tasks, and vendor-specific configuration templates. These two modules collaborate to support various protocols and mixed vendor configurations via a unified graph representation. The Configuration Recommendation Module (CR-Module) utilizes a heterogeneous GNN-based model (HGAT-CR) to perform type-aware reasoning over the CKG and generate top-k candidate parameters. These candidates provide prior knowledge that narrows the search space and improves recommendation accuracy. Finally, they are refined through an LLM-guided optimization mechanism that combines formal verification feedback to produce the final configuration, ensuring maximal intent satisfaction while minimizing side effects. Our evaluation demonstrates that MoCS outperforms existing synthesizers, including NetComplete, INCS, and ConfigReco. In large networks with complex intents, MoCS achieves a high coverage rate (88.23 +/- 1.12%), low redundancy rate (7.89 +/- 1.59%), perfect intent satisfaction rate (1.00 +/- 0.00), and reasonable runtime (143.83 +/- 21.89s). Furthermore, MoCS can synthesize mixed vendor configurations, which current synthesizers cannot handle.

Edge slicing is envisioned to support connected autonomous vehicle (CAV) applications with diverse key performance indicator (KPI) requirements by splitting the shared physical infrastructure into several virtual networks. Unfortunately, existing provisioning approaches struggle to accommodate the spatiotemporal dynamics of CAV traffic, leading to significant violated KPIs or soared resource usage. In this paper, we introduce the transient sharing mechanism among edge slices to obtain reused gains without generating harmful performance interference, in which a slice is allowed to access to the under-utilized reserved resources of other slices but may experience interruptions at any time. Considering the heterogeneity and uncertainty of transient resources, we further develop a two-timescale provisioning scheme. Specifically, slices proactively make reservation decisions based on multi-armed bandit architectures at the beginning of large timescales, while hinging on cost-incentive auction mechanisms selectively preempt transient resources in terms of real-time application demands at each small timescale. With extensive experiments based on real traffic traces, we demonstrate that the proposed scheme can improve 10.43% resource utilization and make slices reduce 42.92% cost than state-of-the-art works, which verifies its high assurance and adaptability.

Cross-tokenizer knowledge distillation, where the teacher and student employ different tokenizers, is becoming increasingly prevalent, yet it poses underexplored challenges: existing methods fail to capture the rich knowledge encoded in teacher logits, as evidenced by the neglect of semantic information, inaccurate and biased logit alignment, and discarding distributional structure—ultimately leading to unfavorable distillation. To address these issues, we propose SeDi, a semantics and distribution-aware knowledge transfer framework tailored for cross-tokenizer distillation. To preserve factual knowledge, SeDi employs bipartite graph-based alignment at the tokenization level and a sliding window re-encoding strategy at the vocabulary level, enabling unbiased transfer of the teacher’s next-token predictions into the student’s vocabulary space. To further retain distributional information, we align the student’s entropy with that of the teacher by incorporating the student’s own logits during training, which helps to mitigate the exposure bias problem. Experiments on ten datasets across three task domains and five different teacher-student model pairs with varying vocabulary sizes demonstrate that SeDi delivers substantial improvements, with gains of up to 19.8%.

Parallelizing CNN inference on heterogeneous edge clusters with data parallelism has gained popularity as a way to meet real-time requirements without sacrificing model accuracy. However, existing algorithms struggle to find optimal parallel granularity for complex CNNS, the structure of which is a directed acyclic graph (DAG) rather than a chain, and the parallel dimension is inflexible. To distribute the workload of modern CNNs on heterogeneous devices is also proven as NP-hard problem. In this paper, we introduce DeepZoning , a versatile and cooperative inference framework that combines both model and data parallelism to accelerate CNN inference. DeepZoning employs two algorithms at different levels: (1) a low-level Adaptive Workload Partition algorithm that uses linear programming and takes spatial and channel dimensions into optimization during the search for feature map distribution on heterogeneous devices, and (2) a high-level Model Partition algorithm that finds the optimal model granularity and organizes complex CNNs into sequential zones to balance communication and computation during execution. Our experimental evaluations show that DeepZoning is effective, achieving up to a 3.02 × speed improvement on our experimental prototype compared to state-of-the-art algorithms.

Federated Learning (FL) enables devices to use data locally for model training and thus has received significant attention for protecting data privacy in the Internet of Vehicles (IoV). However, rational vehicles are reluctant to contribute their data to participate in training without compensation, necessitating the implementation of effective incentive algorithms to motivate vehicles to participate in training. Nevertheless, unlike incentives in other domains, the IoV has the following challenges for the design of incentive systems. First, the large number of vehicles in a dense IoT imposes a huge communication burden and pressure on computational efficiency. Second, road data used by vehicles for training may be affected by factors such as damaged sensors or harsh environments, resulting in changes in data quality. Third, intermittent participation issues are caused by the vehicle’s mobility. To address these issues, we propose DriveFL: A Dynamic Reputation Incentive Mechanism for Federated Learning in Dense Internet of Vehicles. Specifically, we employ gradient compression techniques to reduce communication costs. Subsequently, we design a similarity-based gradient compression quality assessment method capable of evaluating the quality of vehicle data in real time. Then, we develop a dynamic reputation incentive mechanism that quantifies quality assessment records and integrates reverse auction theory, which can attract vehicles with higher data quality from those that intermittently participate in training, thereby enhancing model training quality under constrained communication costs and budget limitations. Theoretical analysis demonstrates that our mechanism satisfies computational efficiency, individual rationality, budget feasibility, and truthfulness. Simulation experiments confirm the effectiveness of our approach.