Pre-trained language models, represented by Bidi-rectional Encoder Representations from Transformers (BERT), show great potential for deep learning-based intrusion detection systems (IDS) due to their strong semantic modeling capability. However, the high cost of training and fine-tuning limits their applicability in large-scale and resource-constrained environments. To address this challenge, we propose a Split-Federated BERT framework with adversarial training for edge intrusion detection. The framework partitions BERT into an Embedding layer deployed at the edge and Transformer and Head layers hosted in the cloud, enabling collaborative training between edge devices and the cloud. At the edge, a conditional generative adversarial network (CGAN) integrated with BERT enhances traffic feature extraction. Guided by BERT, the generator adapts to local traffic distributions, improving sample coverage and feature representation. Edge devices perform local updates to the Embedding layer, while the cloud conducts high-dimensional semantic learning using BERT’s Transformer and Head layers. During federated aggregation, a multi-head attention mechanism is employed in the cloud to differentially weight model updates, ensuring distributional alignment and stable convergence. This design decouples edge-side adversarial enhancement from federated aggregation, reducing both computational and communication overhead. Experimental results on multiple authoritative datasets demonstrate that the proposed method consistently outperforms local deep learning, BERT, federated learning, and split learning baselines in precision, recall, and F1-score, while improving edge computational efficiency and communication cost.

Misinformation detection in social networks faces challenges due to complex semantics, scarcity of labeled data, and rapidly evolving false narratives. To address these issues, we present large language model (LLM)-augmented contrastive learning (LACL), a novel framework that integrates LLMs with contrastive learning (CL) for robust and accurate misinformation detection. We begin with an LLM-driven social media data augmentation strategy, utilizing prompt orchestration to generate diverse yet semantically consistent misinformation samples. These augmented samples are integrated into a CL-based detector, where the semantic richness and diversity introduced by the LLM enhance the CL’s discriminative feature extraction and predictive capability, thus improving generalization beyond the original training data. To align with CL’s discriminative goal, we develop a contrastive loss-aware joint training and fine-tuning approach where CL’s discriminative feature learning actively constrains the LLM’s hallucinations and guides the quality of augmentation. Through this closed-loop optimization, the CL-based detector progressively absorbs latent semantic knowledge from the LLM, effectively overcoming semantic complexity and reducing erroneous generations. Experimental results on four benchmark datasets (Twitter15, Twitter16, Weibo, and PHEME) demonstrate that LACL outperforms mainstream deep learning methods and surpasses approaches that apply commercial LLMs for detection without task-specific adaptation. These results hold consistently across different backbone LLMs (qwen and llama), highlighting LACL’s enhanced robustness, adaptability to varying language contexts, and superior generalization capability.

Joint entity and relation extraction represent a critical task in knowledge representation, but often suffers from the bottleneck of requiring large amounts of labeled data, which is expensive and laborious to obtain. While Semi-Supervised Learning (SSL) offers a way to leverage unlabeled data, traditional methods face limitations in generating high-quality, diverse augmentations for text. This article introduces a novel framework that synergistically combines SSL with Large Language Models (LLMs) to improve joint entity and relation extraction, especially in low-resource settings. Our approach utilizes LLMs to generate semantically coherent and diverse augmented data from unlabeled samples. These augmented samples, along with limited labeled data, are used within an SSL framework employing consistency regularization and pseudo-labeling to train the extraction model. Crucially, the framework incorporates an iterative refinement mechanism where the performance of the SSL component informs the parameter-efficient fine-tuning of the LLM, leading to progressively better data augmentation and model accuracy. We demonstrate through extensive experiments on four benchmark datasets that our proposed method significantly outperforms existing state-of-the-art approaches, particularly when labeled data are scarce. The framework’s design is adaptable and can be integrated with various existing joint extraction models, showcasing its generalizability and practical utility.

This paper presents MobiFormer, a split-federated transfer learning framework with multi-head attention designed for distributed drone Radio Access Network (RAN) slicing. The objective is to optimize slice performance isolation and training costs. Based on a flexible service metric, the problem of maximizing slice performance isolation quality is formulated as a joint optimization of slice windowing and resource allocation. For single drone autonomous operations, we construct an “unconstrained mobility and sustainable fine-tuning” paradigm, enabling drones to adapt previously trained resource slicing models to new environments with the assistance of multiple target-domain terrestrial Base Stations (BSs). This adaptation is facilitated by a Source-free Multi-target-domain Transfer Learning (SMTL) approach, where Transformer-based multi-head attention is employed on drones to integrate fine-tuned models from multiple target-domain BSs. Building on SMTL and continuing its scenario, a Clustered Split Federated Learning (CSFL) approach is developed to support multi-drone collaborative training, where BSs serve as cluster heads to aggregate parameters from member drones. To save energy, part of the onboard models are migrated to BSs while local iterations occur through gradient exchanges. Unlike SMTL, the Transformer is deployed at BSs to enhance the global model's adaptability and generalization. Extensive simulations demonstrate that MobiFormer outperforms benchmark approaches in terms of performance isolation, energy consumption, and online decision-making efficiency in distributed learning settings.

Multi-object detection in autonomous driving faces challenges due to multi-scale entities, diverse streetscapes, and limited computational resources. To address these challenges, we present MT-DyNN, a Multi-Teacher knowledge-distilled Dynamic Neural Network framework for instance-adaptive detection, optimizing detection accuracy and inference cost in autonomous driving. The framework’s student network comprises a customizable multi-branch residual detection network and a lightweight policy network. The former efficiently extracts multi-scale features in parallel without altering receptive fields, while the latter, depending on curriculum learning, captures task-relevant features and dynamically generates routing vectors to guide the activation or deactivation of residual blocks according to image instance complexity. The framework’s teacher network employs a soft-voting strategy to consolidate knowledge from multiple pre-trained teacher models, providing consistent guidance to the student. Within this distillation paradigm, the policy network’s routing search space is gradually refined, and the policy and detection networks are jointly fine-tuned to optimize the alignment between routing decisions and feature extraction. Experimental results on CIFAR and ImageNet demonstrate that compared to early exiting and stochastic depth methods, MT-DyNN achieves higher accuracy at the same inference cost and reduces the cost by 50% and 59% at comparable accuracy levels. The generated routing maintains channel sparsity across diverse scenarios.

Deep reinforcement learning (DRL) has become prevalent for decision-making task assignments in mobile crowdsensing (MCS). However, when facing sensing scenarios with varying numbers of workers or task attributes, existing DRL-based task assignment schemes fail to generate matching policies continuously and are susceptible to environmental fluctuations. To overcome these issues, a twin-delayed deep stochastic policy gradient (TDDS) approach is presented for balanced and low-latency MCS task decomposition and parallel subtask allocation. A masked attention mechanism is incorporated into the policy network to enable TDDS to adapt to task-attribute and subtask variations. To enhance environmental adaptability, an off-policy DRL algorithm incorporating experience replay is developed to eliminate sample correlation during training. Gumbel-Softmax sampling is integrated into the twin-delayed deep deterministic policy gradient (TD3) to support discrete action space decisions and a customized reward strategy to reduce task completion delay and balance workloads. Extensive simulation results confirm that the proposed scheme outperforms mainstream DRL baselines in terms of environmental adaptability, task completion delay, and workload balancing.

A slicing-based collaborative task offloading framework for space-air-ground integrated vehicular networks is proposed in this study, which can provide differentiated quality-of-service (QoS) guarantees for task offloading for high-speed vehicles while maximizing the number of completed tasks. A service-oriented radio access network (RAN) slicing framework is presented that supports slicing window adaptation, spectrum and computing resource orchestration, and collaboration among heterogeneous base stations. Based on the queuing model, the collaborative decision-making of RAN slicing and task offloading is modeled as a problem of maximizing the number of long-term task completions, which consists of three subproblems-slicing window division, resource slicing, and task scheduling-which are solved by a multi-access edge computing (MEC)-enabled controller, forming a closed loop with the slicing window as the period. When a new slicing window arrives, the controller determines its duration according to task traffic fluctuations and allocates resources to RAN slices through an optimization method. A double deep Q-learning network (DDQN)-based algorithm is developed for scheduling workflow on small time scales within a slicing window. Simulation results demonstrate that the proposed scheme performs better than existing approaches in terms of adaptability, task completion rate, and control overhead.

With advancements in cellular vehicle-to-everything (C- V2X) and drone manufacturing technologies, integrating drone-small-cells (DSCs) into terrestrial cellular networks is a promising solution to enabling diversified vehicle applications. In this paper, a multi-DSC-assisted dynamic spectrum management framework is presented to maximize the network utility under quality-of-service (QoS) constraints in 5G and beyond cellular vehicular networks. The network utility maximization problem is formulated as mixed-integer nonlinear programming regarding association patterns between vehicles and base stations (BSs) and spectrum partitioning among heterogeneous BSs. For mathe-matical tractability, the joint optimization problem for spectrum partitioning and vehicle- DSC associations is transformed as a biconcave optimization problem. An alternate search algorithm is then designed to determine vehicle association patterns and spec-trum slicing ratios. Our simulation demonstrates that compared with state-of-the-art methods, the proposed scheme achieves a significant performance improvement in network throughput and spectrum utilization.

Recent years, many applications have been driven advances by the use of Machine Learning (ML). Nowadays, it is common to see industrial-strength machine learning jobs that involve millions of model parameters, terabytes of training data, and weeks of training. Good efficiency, i.e., fast completion time of running a specific ML training job, therefore, is a key feature of a successful ML system. While the completion time of a long-running ML job is determined by the time required to reach model convergence, that is also largely influenced by the values of various system settings. In this paper, we contribute techniques towards building self-tuning parameter servers. Parameter Server (PS) is a popular system architecture for large-scale machine learning systems; and by self-tuning we mean while a long-running ML job is iteratively training the expert-suggested model, the system is also iteratively learning which system setting is more efficient for that job and applies it online. Our techniques are general enough to various PS-style ML systems. Experiments on TensorFlow show that our techniques can reduce the completion times of a variety of long-running TensorFlow jobs from 1.4× to 18×.

Incorporating link correlation awareness into wireless network protocols to facilitate data transmission is an important research issue. In this paper, we focus on link correlation based security threat and countermeasure in wireless networks. By taking advantage of the vulnerability of beacon-based link correlation measurement and the blind spot of malicious node detection mechanisms, we design a new type of link correlation spoofing attack (LCSA) to decrease protocol performance by distorting link correlation information while escaping the tracking of any watchdog and trust systems. Typical cases are analyzed to quantify how the LCSA covertly weakens protocol performance. We also propose beacon-trap (BT), a countermeasure embedded in the beacon-based link condition measurement protocol. Using link diversity as a cover, BT sets traps in the beacon sending sequence to ambush malicious nodes that launch LCSAs without extra control overhead. The performance of BT is not affected by changes in the size of a network or the distribution of nodes. Numerical results demonstrate the superiority and effectiveness of BT against LCSAs in terms of malicious node detection success rate and speed under different parameter settings.