Real-time applications in the key areas of healthcare, the military, and industry require high reliability, throughput, and low latency, often achieved through costly dedicated links and proprietary connectivity solutions. Due to the trend toward shifting service control to data centers, this approach cannot be maintained. As a cost-effective alternative, multi-channel transmission based on recently developed multipath protocols, such as Multipath QUIC, can be used. A multipath protocol distributes the data stream across a few channels running on different physical interfaces, which is managed by a scheduling component. However, standard schedulers tend to prioritize throughput over delay-related factors, making them unsuitable for real-time applications. This work introduces a new path scheduler explicitly designed for delay-sensitive, responsive applications. Reducing the effects of delay variation and losses allows for fast reconstruction of the data stream at the receiver without sacrificing throughput. The design is grounded in formal mathematical principles and has been thoroughly tested on the open Internet against competing solutions.
The rapid growth of encrypted traffic and the lack of high-quality labeled samples pose severe challenges in real-world scenarios. These factors undermine the effectiveness of traditional supervised learning methods for malicious traffic detection. While existing self-supervised learning approaches alleviate data dependency, they primarily rely on single pre-training tasks. Consequently, they struggle to simultaneously capture local structural features and global temporal semantics, leading to limited generalization in few-shot scenarios. To address these challenges, we propose a novel dual-task self-supervised learning framework for few-shot encrypted malicious traffic detection (DSET). First, we construct a CNN–Transformer encoder that processes direction-aggregated burst packet length sequences, designed to model local structural patterns and long-range temporal dependencies for representation learning. Furthermore, we design a dual-task pre-training mechanism that synergizes masked sequence modeling and flow-level contrastive learning. This mechanism integrates local reconstruction constraints with global discriminative consistency over large-scale unlabeled traffic to enforce robust cross-scale semantic alignment. Finally, we devise a parameter-efficient fine-tuning strategy where the backbone remains frozen and only lightweight adapter modules are updated, ensuring stable knowledge transfer with limited labeled samples. Extensive experiments on three public datasets demonstrate that DSET achieves an average F1 score of 94.52% and an AUC of 96.26% in few-shot settings, validating its effectiveness and robustness in detecting unknown attacks.
As IoT devices are widely deployed in various industries, the security threats they confront are becoming increasingly severe, especially in multi-source heterogeneous data environments. Furthermore, the generalization of intrusion detection models poses significant challenges. To address the issues of accuracy degradation and model complexity in cross-domain scenarios, we propose an intrusion detection algorithm based on lightweight deep learning and knowledge transfer-aided genetic optimization, named LightID. It integrates feature alignment techniques and model optimization strategies to build a unified framework that includes data preprocessing, dimensionality reduction, feature space alignment, data transformation, model training, and hyperparameter optimization. To enhance the model’s adaptability and deployment efficiency, we design various lightweight neural networks and perform automated hyperparameter optimization via a genetic algorithm with knowledge transfer. Through two series of experiments conducted on multiple datasets, the proposed algorithm is demonstrated to achieve excellent performance and outperform most peers in both single-domain and cross-domain scenarios. Specifically, in single-domain experiments, it achieves more than 95% detection accuracy in both binary- and multi-classification experiments on all datasets. In cross-domain experiments, the image-based variant achieves detection accuracies above 74% in all cases, with the highest reaching 99.77%, while the CSV-based variant achieves accuracies up to 92.8% in three of six scenarios. These results demonstrate that the proposed algorithm significantly improves cross-domain generalizability and the feasibility of model deployment while obtaining high detection accuracy.
In vehicular networks, each vehicle periodically broadcasts messages via a shared channel to track the movements of surrounding vehicles and detect collision risks. Under heavy traffic conditions, numerous vehicles compete for the limited channel to broadcast their packets, resulting in channel congestion. Existing channel congestion control studies limit the channel load for individual vehicles based on their application requirements. They model application requirements deterministically to enhance safety awareness. However, in high-mobility vehicular networks where traffic conditions and network environments are inherently dynamic, these requirements become probabilistic rather than deterministic. This paper proposes a situation-aware channel congestion control method for vehicular networks, comprising two modules: an assessment module and a decision-making module. The assessment module employs a probabilistic approach—Bayesian network—to evaluate vehicle collision risks due to packet losses. Its probabilistic nature enables robust assessments of potential threats under dynamic and uncertain vehicular communication environments. In the decision-making module, we introduce a distributed channel resource allocation strategy that jointly allocates transmission power and beacon rate for each vehicle in real-time, based on their collision risk levels and real-time channel state information. This approach not only alleviates channel congestion but also enhances vehicles’ environmental perception capabilities. In a high-density traffic environment, compared to existing congestion control methods, our strategy reduces the proportion of medium-risk and high-risk vehicles by 18.01% and 55.41%, respectively.
With the rapid development of information technology, the explosive growth of data has far exceeded the processing capabilities of traditional computing systems. This has given rise to new computing paradigms based on storage and computation platforms such as cloud computing. However, outsourcing massive data to untrusted third-party storage providers poses unprecedented security risks to the data. To ensure the integrity of cloud-stored data, researchers have proposed a variety of data integrity auditing(DIA) schemes. Therefore, we review the existing literature to provide interested readers with a solid and insightful overview of current research advancements. From a system security perspective, we summarize the potential threats faced by DIA systems and propose a structured set of security requirements. In our survey, we also focus on DIA in edge computing storage scenarios, which has often been overlooked in prior studies. This paper introduces an innovative scenario-oriented classification framework that categorizes existing schemes into five representative scenarios based on the interplay between user modes and storage architectures. We systematically analyze representative schemes in each category and conduct detailed comparisons of security, efficiency and functionality. Furthermore, the survey provides a systematic analysis of the key technical approaches employed to achieve various desired properties, summarizing their design rationale, efficiency trade-offs, and supported functionalities. In addition, we conduct a comprehensive performance evaluation and comparative analysis of representative DIA schemes from both theoretical and experimental perspectives. Finally, we identify open challenges and summarize future research directions.
