With the rapid development of cloud computing, network architectures are moving towards distributed computing, which performs data processing at edge nodes to reduce latency, enabling more efficient and scalable network services. Nevertheless, this shift introduces significant security challenges due to the heterogeneity of communications protocols and the vulnerabilities of edge devices. To effectively secure these distributed networks, it is essential to perform multiple traffic analysis tasks, e.g. Network Intrusion Detection, Encrypted Traffic Classification, and Application Traffic Classification. However, existing methods have limited generic feature extraction and require the deployment of multiple models to solve multiple tasks, which exceeds the resource capacity of edge nodes. To address these challenges, we introduce a Lightweight Multitask Traffic Analysis Framework LiMTa, which novelly proposes a traffic pre-training method, FreqRec, and a lightweight multi-task model fine-tune method, MT-Adapter. FreqRec enables high-level semantic feature extraction by reconstructing the frequency features of traffic samples, and MT-Adapter efficiently performs multiple tasks by computing the pre-trained model only once. Experimental results demonstrate that our approach achieves state-of-the-art (SOTA) performance on six traffic analysis tasks. Moreover, the MT-Adapter module only fine-tunes a small number of parameters, accounting for only 6.37% of the pre-trained model’s parameters, and achieves the same result as the full fine-tuning. Compared to full fine-tuning, LiMTa reduces the time cost by 50.9% and the space cost by 57.4% in six edge traffic analysis tasks.

The combination of self-supervised learning and adversarial training (AT) can significantly improve the adversarial robustness of self-supervised models. However, the robustness of self-supervised adversarial training (self-AT) still lags behind that of state-of-the-art (SOTA) supervised AT (sup-AT), even though the performance of current self-supervised learning models has already matched or even surpassed that of SOTA supervised learning models. This issue raises concerns about the secure application of self-supervised learning models. The inclusion of adversarial training turns self-AT into a challenging joint optimization problem, and recent studies have shown that the data augmentation methods necessary for constructing positive pairs in self-supervised learning negatively impact the robustness improvement in self-AT. Inspired by this, we propose 3SAT, a simple self-supervised adversarial training framework. 3SAT conducts adversarial training on original, unaugmented samples, reducing the difficulty of optimizing the adversarial training subproblem and fundamentally eliminating the negative impact of data augmentation on robustness improvement. Additionally, 3SAT introduces a dynamic training objective scheduling strategy to address the issue of model training collapse during the joint optimization process when using original samples directly. 3SAT is not only structurally simple and computationally efficient, reducing self-AT training time by half, but it also improves the SOTA self-AT robustness accuracy by 16.19\% and standard accuracy by 11.41\% under Auto-Attack on the CIFAR-10 dataset. Even more impressively, 3SAT surpasses the SOTA sup-AT method in robust accuracy by a significant margin of 11.25\%. This marks the first time that self-AT has outperformed SOTA sup-AT in robustness, indicating that self-AT is a superior method for improving model robustness.

Self-supervised learning (SSL) is a technique used to learn feature representations from unlabeled data. However, existing SSL frameworks either rely too heavily on domain knowledge due to their design based on feature invariance, leading to a lack of domain transferability, or they are based on autoencoder designs, which generate features with redundant low-level semantics, resulting in suboptimal model representations. In this work, we introduce a novel Domain Agnostic Self-Supervised Learning framework called DASSL, which learns superior high-level feature representations of samples by reconstructing the samples’ missing information in the representation space. DASSL does not require additional domain priors, and compared to successful SSL methods, DASSL achieves competitive representation quality. Moreover, when DASSL incorporates domain-related data augmentation techniques, it outperforms successful methods across multiple datasets and evaluation protocols.

The emergence of Satellite-Terrestrial Integrated Networks (STIN) has significantly expanded terrestrial network coverage but introduced new security threats. Current Intrusion Detection Systems (IDSs) for STIN mostly consider the distributed nature of satellites, overlooking the computational limits of single satellites and the effect of satellite mobility on IDS generalization, where accuracy and adaptability may drop in dynamic environments. To address this, we propose an unsupervised IDS for STIN based on Federated Learning (FL) named STIN-IDS. We deploy IDS in a cross-layer distributed manner, distributing data processing tasks across multiple Low Earth Orbit (LEO) satellites, while Geostationary Earth Orbit (GEO) satellites act as FL clients responsible for anomaly detection, thereby alleviating the computational load on single satellites. Furthermore, to address changes in user regions and traffic patterns due to satellite mobility, FL clients use dynamic data from different LEO regions for collaborative training, improving adaptability to dynamic environments. Experiments across four datasets with varying network conditions show that STIN-IDS achieves strong generalization and outperforms similar methods.

Recently, Deep Learning (DL) models have been increasingly deployed on end-user devices as On-Device AI, offering improved efficiency and privacy. However, this deployment trend poses more serious Intellectual Property (IP) risks, as models are distributed on numerous local devices, making them vulnerable to theft and redistribution. Most existing ownership protection solutions (e.g., backdoor-based watermarking) are designed for cloud-based AI-as-a-Service (AIaaS) and are not directly applicable to large-scale distribution scenarios, where each user-specific model instance must carry a unique watermark. These methods typically embed a fixed watermark, and modifying the embedded watermark requires retraining the model. To address these challenges, we propose Hot-Swap MarkBoard, an efficient watermarking method. It encodes user-specific n-bit binary signatures by independently embedding multiple watermarks into a multi-branch Low-Rank Adaptation (LoRA) module, enabling efficient watermark customization without retraining through branch swapping. A parameter obfuscation mechanism further entangles the watermark weights with those of the base model, preventing removal without degrading model performance. The method supports black-box verification and is compatible with various model architectures and DL tasks, including classification, image generation, and text generation. Extensive experiments across three types of tasks and six backbone models demonstrate our method's superior efficiency and adaptability compared to existing approaches, achieving 100% verification accuracy.

Supervised quantum learning methods face notable limitations in dynamic, real-world environments due to their reliance on static labels and limited adaptability. To address these challenges, we propose a novel online learning framework — Quantum Contextual Bandit (QCB) — that integrates quantum neural networks (QNNs) with contextual bandit (CB) algorithms. The QCB framework enables adaptive decision-making by incorporating bandit-based exploration into QNN training, making it particularly suitable for applications such as recommender systems. To mitigate the adverse effects of quantum noise—including depolarizing, Pauli, and shot noise, the framework leverages a gradient-free optimization approach, enhancing robustness and convergence stability. Experimental results on various datasets demonstrate that QCB consistently outperforms traditional QNN training methods with identical circuit architectures. Notably, the model achieves over 99% accuracy under ideal conditions and sustains high performance under noisy quantum environments. These results underscore the potential of QCB as a scalable, noise-resilient solution for adaptive learning in quantum machine learning systems.

The signaling collected in mobile communication networks can intuitively display the operational status of the system, which can use to locate faults. This paper proposes a novel signaling-based end-to-end fine-grained troubleshooting (simFGT) method for 6G NTN networks. First, the signaling collected from the core network is analyzed to extract multidimensional KPIs and attribute information. Second, a root cause localization algorithm is employed for fine-grained fault localization. Third, a lightweight data-driven ensemble learning method is adopted, with the abnormal KPI of the localized root cause node as inputs, and precise fault classification is achieved through data-driven weight optimization. Experiments results show that the proposed lightweight SimFGT method achieves best balance between precision and recall, resulting in highest F1 score, outperforming current state-of-the-art solutions.

Recently, Self-Supervised Learning (SSL) has achieved great success in various famous applications e.g., BERT and ChatGPT. However, when applying SSL to safety-critical downstream tasks, such as self-driving cars, potential adversarial attacks can completely change the final decisions and thus lead to serious security issues. To overcome this issue, existing methods combine adversarial training with pre-training to improve the adversarial robustness of SSL. However, combining these two computationally complex processes may largely amplify the computation cost. Moreover, whether performing adversarial training in pre-training or fine-tuning, current methods may degrade the accuracy due to the famous catastrophic forgetting problem. The computation cost of current adversarial training methods based on full parameter updating is still high even in the fine-tuning stage.To address the above challenges, we propose an effective robust fine-tuning framework for SSL based on Low-Rank Adaptation (LoRA), named LoFT. First, LoFT performs adversarial training in the fine-tuning stage rather than in the pre-training stage. Second, LoFT innovatively and elaborately integrates LoRA into adversarial training to avoid the catastrophic forgetting problem. Third, LoFT exploits a low-rank matrix in LoRA which enables efficient fine-tuning by updating only a small set of parameters, which contains only 1%-5% of the parameters of the pre-trained model. The whole pre-training and fine-tuning stages take only 9.44 hours, which reduces training time by 3× over the current SOTA method. Furthermore, compared with existing SOTA robust pre-training methods for SSL, LoFT improves accuracy by 5.97% (77.41%⇒83.38%) and robustness by 13% (45.04%⇒58.44%) on the CIFAR-10 dataset.