Speech spoofing detection (SSD) requires capturing complex dependencies across temporal, spectral, and both short- and long-term artifacts. While Mamba-based models have shown promise in SSD, their typically simple fusion strategies limit the joint modeling of diverse artifacts. In this work, we propose a novel Tri-Attention Fusion module that progressively integrates the bidirectional dual-branch outputs of BiMamba-ST to produce highly discriminative features. Within this module, Local-domain Attention adaptively fuses bidirectional features along the channel dimension, Cross-domain Attention enables effective spectro-temporal interactions, and Global Representation Pooling unifies the stage-wise representations. We further demonstrate the module’s versatility across different frontends, including end-to-end and pre-trained systems. Experiments on ASVspoof 2019 LA, 2021 LA, 2021 DF, and In-the-Wild datasets show that the proposed countermeasure achieves performance comparable to or exceeding current state-of-the-art methods.

User-generated content (UGC) facilitates cross-platform user identity linkage (UIL), increasing the risk of privacy leakage. Existing anti-UIL methods overlook adversarial strategies based on UGCs. To address this gap, we propose the Adversarial Multimodal User-Generated Content Generation (AdvUGCG) method, which includes adversarial example generation for user-generated text and user-generated images. First, we design a multimodal confusion attack module for adversarial text generation. The module splits UGC text into functional words, entity words, and general words, and generates character-level and word-level adversarial text through multimodal confusion. Next, we propose a text-guided universal adversarial perturbation (UAP) generation module for adversarial image generation. In this module, we fine-tune the UAP generator using social text as cross-modal guidance. Moreover, we propose two novel data augmentation methods to enhance the transferability of the UAP. We conduct experiments on the TWFQ and TWIN real-world social datasets targeting the EUIL and AHG-Net UIL models. Results show that the processed data reduced the accuracy of these models to 55.88 % and 86.75 %, respectively, significantly decreasing the success rate of UIL and providing an effective method to protect the privacy of social network users. The code is available at: https://github.com/JessieGaoo/AdvUGCGhttps://gith ub.com/JessieGaoo/AdvUGCG.

Software vulnerabilities have increased sharply, underscoring the growing urgency for effective detection methods. Although large language model (LLM) based methods have shown promise in this task, current state-of-the-art LLM approaches struggle with functions that have long contexts. In this paper, we propose CTX-Coder, a context-enhanced vulnerability detection framework that enables LLMs to selectively focus on relevant contextual functions. To achieve this, we represent the contextual functions as embeddings and integrate them with the target code via cross-attention, thereby enhancing the model's ability to capture contextual information. Furthermore, to equip the model with the ability to recognize these embedding features, we propose a two-stage pretraining pipeline. We also introduce a new dataset, CTX-VUL, which addresses the limitations of existing datasets that either lack contextual information for vulnerable functions or are not publicly available. Extensive experiments demonstrate that CTX-Coder (10B) significantly outperforms baseline models with even larger parameters, such as Qwen2.5-14B and SecGPT. As the input code length increases, CTX-Coder’s F1 score drops by only 5.01%, while other models degrade by 25% to 41.5%, showing strong robustness to long-context scenarios and the effectiveness of our design.

Recent advances in Large Language Models (LLMs) have shown considerable potential for automating source-level vulnerability detection. However, existing supervised fine-tuning approaches for LLMs are constrained by the scarcity of high-quality SFT datasets, and therefore typically produce only binary predictions indicating the presence or absence of vulnerabilities. As a result, they fail to fully leverage a key advantage of LLMs: providing developers with informative reasoning about vulnerability causes and actionable remediation guidance. To enable developer-friendly vulnerability analysis under limited labeled data, we propose MCTS-VUL, a teacher–student self-training framework that integrates LLMs with Monte Carlo Tree Search (MCTS) to generate structured reasoning traces for vulnerability detection. MCTS-VUL employs a value model to guide the search process and identify high-quality reasoning paths. Experimental results demonstrate that our approach not only generates developer-oriented reasoning traces, but also achieves competitive or state-of-the-art performance across three public datasets. These results further indicate that the generated reasoning traces effectively facilitate LLMs in learning the underlying causes of vulnerabilities.

Most current deepfake detection studies lack curriculum learning strategies that adapt to the intrinsic difficulty of individual samples, which prevents models from efficiently focusing on key forgery cues and ultimately limits cross-domain generalization. To address this, we propose Dynamic Frequency-domain Curriculum Learning (DFCL), a framework that guides an easy-to-hard progression via the joint optimization of a Frequency-domain Difficulty Score (FDS) and a Dynamic–Static Curriculum Scheduling (DSCS) policy. Concretely, FDS is constructed from the Fourier Correlation Variance (FCV) computed over block-wise Fourier transforms and is used to quantify a sample’s intrinsic difficulty; the dynamic loss in DSCS reflects the model’s on-the-fly discriminative capability. The two are fused into a hybrid difficulty score by a training-epoch–dependent weighting function γ(t). Based on this score, training proceeds in three stages: An easy-to-hard curriculum: high γ(t) prioritizes low-FDS samples, decaying γ(t) introduces harder ones, and γ(t) = 0 fine-tunes on all data. Extensive experiments on Celeb-DF, DFDCP, and DFDC demonstrate that DFCL markedly improves the detection performance of mainstream backbones such as ResNet, EfficientNet, and Xception. With a ViT (CLIP) backbone, DFCL achieves state-of-the-art performance. In addition, DFCL reduces training time by roughly 40% while maintaining accuracy.

With the popularity of social networks, social bots are increasingly interfering with human social activities, which have negatively impacted social network ecology and user experience. How to detect social bots effectively has become a challenging task. Most of existing social bot detection methods are based on the supervised learning mechanism, but the labelled data collection is resource-intensive. In this paper, a Multi-stage Selftraining Social Bot detection method based on a Graph Neural Network (MSSBot) is proposed to improve the performance of social bot detection with a small amount of labelled data. MSSBot exploits pseudo-labels to expand the dataset and trains a Relational Graph Convolutional Network in multiple stages, while uncertaintyaware and deep clustering are used to reduce the noise present in the pseudo-labels acquisition process. The experiments were conducted on a real-world social network dataset Twibot-20, a small high-quality dataset focused on social bot behavior patterns Cresci-2015, and a large-scale graph-based dataset Twibot-22 with high annotation quality. The experimental results show that our method achieves accuracies of 0.9259, 0.9979, and 0.8358 on the Twibot-20, Cresci-2015, and Twibot-22 datasets, respectively, outperforming the state-of-the-art methods, with improvements of 5.68 %, 2.64 %, and 3.92 % over the existing state-of-the-art methods.

Low rank adaptation (LoRA) methods have demonstrated strong capabilities in efficiently fine-tuning large models. However, existing LoRA-based approaches typically require manually setting the scaling factor, a process that involves extensive search efforts to find optimal values. To address this challenge, we first develop data-driven heuristic methods that automatically determine layer-wise scaling factors through either activation pattern analysis during forward propagation or gradient behavior monitoring during backward updates. However,their practical performance remains unsatisfactory in applications. Building upon these theoretical foundations, we present MSLoRA, a novel framework that reformulates scaling factor determination as a dynamic optimization problem in parameter-efficient fine-tuning. Our approach innovatively models scaling factors as self-adaptive meta-parameters whose optimal values emerge organically through the interplay between transformer architecture hierarchies and task-specific learning objectives. Extensive experiments conducted across both natural language understanding and generative tasks reveal that MSLoRA consistently outperforms baseline models. This highlights the effectiveness of MSLoRA's dynamic, layer-specific adjustment mechanism in capturing the complex nature of task-specific activation patterns, making it a more robust and scalable solution for parameter-efficient fine-tuning of large models.

Despite their potential, the industrial deployment of large language models (LLMs) is constrained by traditional fine-tuning procedures that are both resource-intensive and time-consuming. Low-Rank Adaptation (LoRA) has emerged as a pioneering methodology for addressing these challenges. By integrating low-rank decomposition matrices into network weights to reduce trainable parameters, LoRA effectively accelerates the adaptation process. While research on LoRA primarily focuses on adjusting low-rank matrices, DyLoRA optimizes the rank-setting mechanism to avoid extensive effort in rank size training and searching. However, DyLoRA rank configuration mechanism has its own limitation. First, DyLoRA sets the same rank size for all the low-rank adaptation layers at each time step. Given that layers with different depth contain distinct information, they should have varying rank values to accurately capture their unique characteristics. Second, the truncated phase selected for ordering representation based on nested dropout regulation is only half dynamic, continuously dropping tail units, thereby limiting its ability to access information. In this work, we propose a novel technique, enhanced range adaptation in time and depth aware dynamic LoRA (ERAT-DLoRA) to address these problems. The ERAT-DLoRA method introduces a dynamic range to the truncated phase that makes the truncated phase fully dynamic. Additionally, we design a time and layer-aware dynamic rank to ensure appropriate adjustments at different time steps and layer levels. We evaluate our solution on natural languages understanding and language generation tasks. Extensive evaluation results demonstrate the effectiveness of the proposed method.

Cognitive biases are commonly used by attackers to manipulate users’ psychology in phishing emails. This study systematically analyzes the exploitation of cognitive biases in phishing emails and addresses the following questions: (1) Which cognitive biases are frequently exploited in phishing emails? (2) How are cognitive biases exploited in phishing emails? (3) How effective are cognitive bias features in detecting phishing emails? (4) How can the exploitation of cognitive biases in phishing emails be modelled? To address these questions, this study constructed a cognitive processing model that explains how attackers manipulate users by leveraging cognitive biases at different cognitive stages. By annotating 482 phishing emails, this study identified 10 common types of cognitive biases and developed corresponding detection models to evaluate the effectiveness of these bias features in phishing email detection. The results show that models incorporating cognitive bias features significantly outperform baseline models in terms of accuracy, recall, and F1 score. This study provides crucial theoretical support for future anti-phishing methods, as a deeper understanding of cognitive biases offers key insights for designing more effective detection and prevention strategies.

The emergence of malicious Twitter social bots poses a considerable threat to the security of social networks, and the detection of evolving social bots has become challenging. State-of-the-art detection methods are usually supervised, but the label acquisition process suffers from time-consuming and inaccurate problems in addition to its inability to cope with the challenge of the continuous evolution of social bots. Moreover, structural relationships within social networks are under-explored using current unsupervised methods. To address current challenges, we propose an unsupervised social bot detection method based on deep contrastive graph clustering (BotDCGC). This approach utilises a graph attentional encoder and an inner product decoder to acquire user node embeddings. By incorporating information from both user account features and topological structure, the model employs a contrastive learning technique based on structure to discern node embeddings of distinct classes within the feature space. Subsequently, confident cluster assignments are used as soft labels to guide the embedding process by calculating the similarity between each node and the clustering centre, enabling the joint optimisation of user node embeddings and clustering results. Experiments show that BotDCGC is more effective than the state-of-the-art baselines with an accuracy of 0.8095 in the Twibot-20 dataset and 0.9334 in the Cresci-2015 dataset, and the experimental results indicate the effectiveness of the graph autoencoder module, contrastive learning and deep clustering module.

Malicious social bots pose a serious threat to social network security by spreading false information and guiding bad opinions in social networks. The singularity and scarcity of single organization data and the high cost of labeling social bots have given rise to the construction of federated models that combine federated learning with social bot detection. In this paper, we first combine the federated learning framework with the Relational Graph Convolutional Neural Network (RGCN) model to achieve federated social bot detection. A class-level cross entropy loss function is applied in the local model training to mitigate the effects of the class imbalance problem in local data. To address the data heterogeneity issue from multiple participants, we optimize the classical federated learning algorithm by applying knowledge distillation methods. Specifically, we adjust the client-side and server-side models separately: training a global generator to generate pseudo-samples based on the local data distribution knowledge to correct the optimization direction of client-side classification models, and integrating client-side classification models’ knowledge on the server side to guide the training of the global classification model. We conduct extensive experiments on widely used datasets, and the results demonstrate the effectiveness of our approach in social bot detection in heterogeneous data scenarios. Compared to baseline methods, our approach achieves a nearly 3–10% improvement in detection accuracy when the data heterogeneity is larger. Additionally, our method achieves the specified accuracy with minimal communication rounds.

Black-box attacks generate adversarial examples by querying the target model and updating the noise according to the feedback. However, the current black-box attack methods require excessive queries to generate adversarial examples, increasing the risk of detection by target defense systems. Furthermore, the current black-box attack methods primarily focus on controlling the magnitude of perturbations while neglecting the impact of perturbation placement on the stealthiness of adversarial examples. To this end, we propose a novel edge noise-constrained black-box attack method using the artificial fish swarm algorithm (EFSAttack). EFSAttack introduces the concept of edge noise constraint to indicate the low-frequency region of the image where perturbations are added and employs edge noise constraint to improve the population initialization and population evolution process. The experiments on CIFAR-10 and MNIST show notable improvements in the success rates, query efficiency, and adversarial example invisibility.

User identity linkage (UIL) refers to identifying user accounts belonging to the same identity across different social media platforms. Most of the current research is based on text analysis, which fails to fully explore the rich image resources generated by users, and the existing attempts touch on the multimodal domain, but still face the challenge of semantic differences between text and images. Given this, we investigate the UIL task across different social media platforms based on multimodal user-generated contents (UGCs). We innovatively introduce the efficient user identity linkage via aligned multi-modal features and temporal correlation (EUIL) approach. The method first generates captions for user-posted images with the BLIP model, alleviating the problem of missing textual information. Subsequently, we extract aligned text and image features with the CLIP model, which closely aligns the two modalities and significantly reduces the semantic gap. Accordingly, we construct a set of adapter modules to integrate the multimodal features. Furthermore, we design a temporal weight assignment mechanism to incorporate the temporal dimension of user behavior. We evaluate the proposed scheme on the real-world social dataset TWIN, and the results show that our method reaches 86.39% accuracy, which demonstrates the excellence in handling multimodal data, and provides strong algorithmic support for UIL.

The number of social bots in OSNs(Online Social Networks) ,has increased significantly, which brings great risks to social network security. Aiming at the problem of insufficient emoji usage in the existing social bot detection, this paper proposes a social account representation method combining emoji information and is used for social bot detection. The emojiswitch library was used to replace the text of emoji, and the word embedding model RoBERTa was used to obtain the tweet features. User description, user tweets, user numerical properties, user categorical properties and user relationship are combined to realize the complete emoji of social accounts. The effectiveness of the proposed method is verified based on the public data set.

Deep Learning algorithms have achieved state-of-the-art performance in various important tasks. However, recent studies have found that an elaborate perturbation may cause a network to misclassify, which is known as an adversarial attack. Based on current research, it is suggested that adversarial examples cannot be eliminated completely. Consequently, it is always possible to determine an attack that is effective against a defense model. We render existing adversarial examples invalid by altering the classification boundaries. Meanwhile, for valid adversarial examples generated against the defense model, the adversarial perturbations are increased so that they can be distinguished by the human eye. This paper proposes a method for implementing the abovementioned concepts through color space transformation. Experiments on CIFAR-10, CIFAR-100, and Mini-ImageNet demonstrate the effectiveness and versatility of our defense method. To the best of our knowledge, this is the first defense model based on the amplification of adversarial perturbations.

Residual Network (ResNet) is a distinguished network structure in deep learning, and its layers can be profound. We theoretically explore the mathematics characteristics of the ResNet, in particular, to pay attention to the gradient information, which is a powerful and straightforward mathematical tool for analysing the properties of ResNet, such as the gradients of the loss function with respect to the input and the weight parameters and the gradient of the entry of the logits with respect to the input. A theorem about the explicit matrix expression of gradients in Resnet is given in this work. A rigorous mathematical and logical derivation of the theorem is obtained in detail by the matrix derivative definition and matrix differentiation. We further provide explicit matrix expressions of some deep learning algorithms in ResNet, including backpropagation, gradient-based adversarial attacks, and gradient-based saliency maps. Furthermore, the reasons why the ResNet network works are analysed. Finally, experimental results are provided to verify the correctness and efficiency of the proposed theorem.

文本情绪分析在日益发展的互联网时代极为重要，其在用户评价和舆情分析等领域具有广泛应用。但当前的研究方法很少综合考虑多个情绪维度做出判断，缺乏对情感间共现关系的理解。针对该问题，本文提出了一种融合对比损失的层级式Transformer模型，主要的改进思路是在文本分类训练过程中引入更丰富信息。实验结果表明，在情绪分布预测任务中，该模型有效增强了编码器的建模能力，在无需其它适配下游任务的预处理调整的条件下，超越了当前最优的模型表现，验证了模型的有效性。

Popular topics in the field of countermeasures include feature engineering and neural-network-based models, which involve neural network architectures and loss criteria. This study focuses on Res2Net and its variant models to examine the impact of model generalization on countermeasure performance in the ASVspoof 2019 logical access and physical access scenarios. Results reveal that while Res2Net exhibits superior generalization compared to its variants, the most effective countermeasure combines both feature engineering and model optimization. The proposed dynamic modulated-Res2Net utilizes channel-wise soft attention to recalibrate feature maps, offering adaptive adjustments to spoofing cues of varying scales. Evaluation on the logical access dataset demonstrates dynamic modulated-Res2Net's relative improvement of over 38% compared to Res2Net. Furthermore, we exploit low-frequency features and combine them with dynamic modulated-Res2Net to achieve in an equal error rate of 1.21% under logical access and 0.41% under physical access, establishing our proposed dynamic modulated-Res2Net as one of the top-performing single systems. Additionally, we compare the best countermeasures in different scenarios, highlighting the ongoing challenge of achieving generalization.

针对网络安全课程知识点分散且难度大,所学知识不能系统理解、无法举一反三的问题,提出将成语及"三十六计"为代表的优秀传统文化融入教学环节,介绍具体教学应用过程,最后以实际教学数据说明教学应用效果.

Social engineering attacks are considered one of the most hazardous cyberattacks in cybersecurity, as human vulnerabilities are often the weakest link in the entire network. Such vulnerabilities are becoming increasingly susceptible to network security risks. Addressing the social engineering attack defense problem has been the focus of many studies. However, two main challenges hinder its successful resolution. Firstly, the vulnerabilities in social engineering attacks are unique due to multistage attacks, leading to incorrect social engineering defense strategies. Secondly, social engineering attacks are real-time, and the defense strategy algorithms based on gaming or reinforcement learning are too complex to make rapid decisions. This paper proposes a multiattribute quantitative incentive method based on human vulnerability and an improved Q-learning (IQL) reinforcement learning method on human vulnerability attributes. The proposed algorithm aims to address the two main challenges in social engineering attack defense by using a multiattribute incentive method based on human vulnerability to determine the optimal defense strategy. Furthermore, the IQL reinforcement learning method facilitates rapid decision-making during real-time attacks. The experimental results demonstrate that the proposed algorithm outperforms the traditional Q-learning (QL) and deep Q-network (DQN) approaches in terms of time efficiency, taking 9.1% and 19.4% less time, respectively. Moreover, the proposed algorithm effectively addresses the non-uniformity of vulnerabilities in social engineering attacks and provides a reliable defense strategy based on human vulnerability attributes. This study contributes to advancing social engineering attack defense by introducing an effective and efficient method for addressing the vulnerabilities of human factors in the cybersecurity domain.