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.

Driven by advances in GANs and diffusion models, deepfake content has reached an unprecedented level of photorealism, causing detectors to deteriorate once they leave their training domain. Most prior studies adopt CLIP as the backbone of an image-level binary classifier, yet overlook CLIP’s core strength: text-to-image semantic alignment. Moreover, captions generated by CLIP-CAP lack sufficient high-level semantics to distinguish between authentic and manipulated faces. Deepfake generators often fail to maintain semantic coherence, resulting in contradictions that traditional visual models cannot capture. Existing approaches also intermingle all samples during training and thus lack a systematic, difficulty-aware curriculum. To bridge these gaps, we introduce Semantic- and Frequency-Enhanced (SAFE) deepfake detection, a two-component framework: 1) Semantic-enhanced multimodal alignment. Authenticity cues are injected into CLIP-CAP captions, and low-rank LoRA fine-tuning is applied to CLIP’s visual branch, yielding dual supervision for text–image alignment and forgery discrimination. 2) Dual-score curriculum learning. Fourier Correlation Variance (FCV) measures local spectral consistency and, combined with the loss value, is transformed into a difficulty score that ranks training samples from easy to hard, reducing training time by 23.3% and enhancing generalization. SAFE attains state-of-the-art performance on several cross-dataset and cross-manipulation benchmarks. Ablation studies confirm that semantic enhancement, LoRA fine-tuning, and dual-score curriculum are complementary, jointly delivering substantial gains in open-set generalization.

Federated clothing recommendation suggests clothing items to customers (users) based on their past purchase behaviors in a privacy-preserving manner. Current federated recommendation systems face two main challenges. The first is limited information exploration. Existing methods mainly rely on item ID-based embedding and ignore multimodal information about items. The second challenge, data heterogeneity, arising from that users perceive the same clothing item differently, leads to performance degradation. We therefore propose a multimodal federated learning framework for personalized clothing recommendation (MMFashion). To address the first challenge, we represent each item using its image and text descriptions, generating rich item representations from which user preferences can be derived. To address the second challenge, we introduce personalized image and text adapters to guide the local model in capturing item attributes that align with user’s interests. Such user-specific item representations effectively capture individual preferences. We also introduce a threshold-based regularization loss to push the learned embeddings of user preferences and dislikes apart. To evaluate the effectiveness of MMFashion, we construct datasets by extracting two clothing types (i.e., long sleeves and outerwear) from a real-world clothing dataset released by a fashion retail company. Experimental results show that MMFashion consistently outperforms existing methods, supporting its effectiveness in addressing the challenges in federated clothing recommendation.

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.

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.

Artificial intelligence is developing rapidly in the direction of intellectualization and humanization. Recent studies have shown the vulnerability of many deep learning models to adversarial examples, but there are fewer studies on adversarial examples attacking facial expression recognition systems. Human–computer interaction requires facial expression recognition, so the security demands of artificial intelligence humanization should be considered. Inspired by facial expression recognition, we want to explore the characteristics of facial expression recognition adversarial examples. In this paper, we are the first to study facial expression adversarial examples (FEAEs) and propose an adversarial attack method on facial expression recognition systems, a novel measurement method on the adversarial hardness of FEAEs, and two evaluation metrics on FEAE transferability. The experimental results illustrate that our approach is superior to other gradient-based attack methods. Finding FEAEs can attack not only facial expression recognition systems but also face recognition systems. The transferability and adversarial hardness of FEAEs can be measured effectively and accurately.

Nearest neighbor (NN) models play an important role in the intrusion detection system (IDS). However, with the advent of the era of big data, the NN model has the disadvantages of low efficiency, noise sensitivity, and high storage requirement. This paper presents a neighbor prototype selection method based on CCHPSO for intrusion detection. In the model, the prototype selection and feature weight adjustment are performed simultaneously and k-nearest neighbor (KNN) is used as the basic classifier. To deal with large-scale optimization problems, a cooperative coevolving algorithm based on hybrid standard particle swarm and binary particle swarm optimization, which employs the divide-and-conquer strategy, is proposed in this paper. Meanwhile, a fitness function based on the accuracy and data reduction rate is defined in the CCHPSO to obtain a set of appropriate prototypes and feature weights. The KDD99 and NSL datasets are used to assess the effectiveness of the method. The empirical results indicate that the data reduction rate of the proposed method is very high, ranging from 82.32% to 92.01%. Compared with all the data used, the proposed method can not only achieve comparable accuracy performance but also save a lot of storage and computing resources.

Machine learning plays an important role in constructing intrusion detection models. However, the information era is an era of data. With the continuous increase in data size and the growth of data dimensions, the ability of a single classifier is becoming limited in predicting samples. In this paper, we present an ensemble method using random subspace in which an extreme learning machine (ELM) is chosen as the base classifier. To optimize the ensemble model, an ensemble pruning method based on the bat algorithm (BA) is proposed. Meanwhile, a fitness function based on the accuracy and diversity of an ensemble is defined in the BA to obtain an improved classifier subset. Three public datasets, the KDD99, NSL and Kyoto datasets, are adopted to assess the robustness of the method. The empirical results indicate that the ensemble method based on random subspace can improve the accuracy and robustness over the use of an individual ELM. The results also show that compared with when all the sub-classifiers are used in the ensemble, the pruning framework can not only achieve comparable or better performance but also save substantial computing resources in an intrusion detection system (IDS).

Android malware detection is a complex and crucial issue. In this paper, we propose a malware detection model using a support vector machine (SVM) method based on feature weights that are computed by information gain (IG) and particle swarm optimization (PSO) algorithms. The IG weights are evaluated based on the relevance between features and class labels, and the PSO weights are adaptively calculated to result in the best fitness (the performance of the SVM classification model). Moreover, to overcome the defects of basic PSO, we propose a new adaptive inertia weight method called fitness-based and chaotic adaptive inertia weight-PSO (FCAIW-PSO) that improves on basic PSO and is based on the fitness and a chaotic term. The goal is to assign suitable weights to the features to ensure the best Android malware detection performance. The results of experiments indicate that the IG weights and PSO weights both improve the performance of SVM and that the performance of the PSO weights is better than that of the IG weights.

Advanced Persistent Threat (APT) imposes increasing threats on cyber security with the developing network attack technologies. APT is a highly interactive, specifically targeted and extremely harmful network-centric attack, which employs various technologies to evade detection during attacks leading to the result that victims will not be aware of attacks until they suffer from tremendous losses. Since command and control (C&C) is an essential component during the lifetime of APT, the detection of it is a practical measure to defend against the APT. In this paper, we analyze the features of C&C in APT and find that the HTTP-based C&C is widely used. Based on the analysis results, we propose a new feature of C&C, i.e., independent access, to characterize the difference between C&C communications and normal HTTP requests. Applying the independent access feature into DNS records, we implement a novel C&C detection method and validate it on public dataset. As a new feature of C&C, its advantages and drawbacks are also analyzed.