To address the issues of utility degradation in deep learning models under differential privacy protection and the gap between theoretical and actual privacy protection effectiveness, this paper proposes a deep learning differential privacy protection scheme based on adaptive clipping. The scheme optimizes the process through a fourstep mechanism: firstly, gradient adaptive clipping controls the gradient magnitude during training by dynamically adjusting the gradient clipping threshold, thereby enabling the control of the magnitude of noise added subsequently; secondly, group label selection identifies the group with the smallest gradient as the privacypreserving object, and more accurate privacy loss can be obtained by training this group; thirdly, optimized privacy loss calculation combines the gaussian mechanism based on subsampling to reduce the computational overhead of model privacy loss calculation; finally, optimized gradient adaptive descent realizes the adaptive descent of gradients by adjusting the conditional smoothing parameter, thus improving the usability of the model. Experiments were conducted on the VGG architecture using the MNIST, CIFAR10, and MedicalMNIST datasets. The results show that the model accuracy rates after training with this scheme are 81.08%, 72.30%, and 67.91% respectively, representing improvements of 15.60%, 10.60%, and 9.71% compared to the traditional DPSGD, and 0.63%, 2.50%, and 4.40% over the widely used Nadam algorithm in recent years. The model training efficiency has been improved by 35.5% and 39.4%, respectively.

Most of the smart contract vulnerability detection methods rely on single mode feature extraction, which leads to the problem of low detection accuracy due to insufficient key feature extraction. This paper proposes a smart contract vulnerability detection method based on multimodal feature fusion. Firstly, the construction of the control flow graph (CFG) is constructed by leveraging the abstract syntax tree (AST) trimmed at the source code layer and the data flow relationship based on the opcode layer, which is imported into the graph attention network (GAT) to extract two types of static features. Secondly, the fuzzing test report generated by echidna, a dynamic detection tool, is used to extract path coverage, state changes and other information to build a graph model, and the dynamic features are extracted by graph neural network (GNN). Finally, the extracted static and dynamic features are fused and input into CNN bilstm att model for vulnerability detection, and relevant experiments are carried out on 47398 smart contracts. Experimental results show that compared with eight mainstream detection methods, such as SmartCheck, Mythril, Oyente, BiGGNN, ASTNN, DRGCN, SVCB and CBGRU, the accuracy, recall and F1 value of this method in reentry vulnerability, timestamp vulnerability, integer overflow vulnerability and Tx.origin vulnerability are increased by 50.26%, 59.54% and 58.40%.

Deep neural networks are highly vulnerable to the threat of backdoor attacks due to their noninterpretability and high dependence on data during training. Although the current mainstream backdoor attack methods generally use fixed trigger design to simplify implementation, these triggers are often significantly different from the training data distribution, resulting in easy detection and identification. To this end, this paper proposes a dynamic invisible backdoor attack method via frequency domain injection: firstly, a generative network is used to generate a specific trigger pattern based on the input samples, and then the highfrequency information of the pattern is injected into the wavelet domain of the samples, ensuring the triggers remain stealthy. Additionally, this paper designs a fair screening strategy to select samples that are more influential to the backdoor model through cosine similarity and Kmeans clustering algorithm. Experimental results show that this method outperforms existing methods (e.g., BadNets, Blend, WaNet, and WABA) in terms of attack success rate and stealthiness, and effectively circumvents a variety of stateoftheart defence mechanisms (e.g., FP, NC, SentiNet, and SCALEUP), providing significant robustness and extensive practical potential.

Aiming at addressing existing issues in current image encryption technologies regarding encryption speed, security, and sensitivity, this paper proposes a novel image encryption method based on a combined chaotic system and fractional numbertheoretic transform. First, a new chaotic structure is proposed by combining two traditional onedimensional mappings to create a fully chaotic mapping. Metrics such as bifurcation diagrams, Lyapunov exponents, and information entropy demonstrate that the proposed chaotic structure exhibits excellent chaotic performance, large parameter space, strong sensitivity, and high randomness. Subsequently, a new image encryption method is developed based on this chaotic mapping and multiparameter fractional number theoretic transform. The hash value of the plaintext image is linked with the parameters of the chaotic system to generate initial chaotic keys and scrambling parameters. A multiparameter fractional number theoretic transform is defined by constructing a number theoretic transform feature vector. The plaintext image undergoes one round of number theoretic transform to obtain an intermediate image, followed by Arnold scrambling to disrupt the image. Finally, another round of numbertheoretic transformation is applied to generate the ciphertext image. Experimental results indicate that the algorithm achieves excellent encryption performance: the pixel change rate (NPCR) and unified average changing intensity (UACI) closely approach their ideal values; the average correlation coefficient of ciphertext images is 0.0018, approaching zero; the normalized entropy of ciphertext images reaches 0.9994, nearing the maximum value of 1. With an average encryption time of 0.273s and decryption time of 0.324s, the method outperforms other comparative schemes in efficiency. It demonstrates robust resistance against common attacks including chosenplaintext attacks, differential attacks, and exhaustive attacks, exhibiting high security and promising application prospects in multimedia security fields.

The increasing complexity of cyber attacks challenges traditional centralized intrusion detection systems. Federated learningbased collaborative intrusion detection enables collaborative modeling and knowledge sharing among multiple nodes without sharing raw data, thereby effectively improving the detection capability for crossdomain and unknown attacks. This paper systematically reviews the research progress of federated learningbased collaborative intrusion detection. Existing methods are classified and analyzed from multiple perspectives, including architectureaware, model adaptation and evolutiondriven, as well as privacy and security enhanced approaches. Commonly used datasets and evaluation metrics are summarized. Finally, the major challenges and future research directions are discussed, providing references for subsequent research in this field.

This paper proposes an edge weightaware graph neural network (EWGNN) model for encrypted traffic classification. By introducing an innovative edgeweighting mechanism, the model effectively leverages graph structural information to distinguish the importance of different edges for classification tasks, thereby enhancing feature extraction capabilities while reducing noise interference. The EWGNN architecture comprises four core components: a dualbranch embedding structure, a GNNbased traffic representation encoder, a crossgating feature interaction mechanism, and an endtoend classification module. Experimental results demonstrate that EWGNN achieves 94.75% accuracy, 95.12% precision, 94.83% recall, 94.97% F1score, and 0.954 AUC on the ISCXVPN dataset, significantly outperforming baseline models. Ablation studies further validate the effectiveness of the edgeweighting mechanism, showing over a 1.5% performance improvement across all metrics when activated. Future work will focus on extending application scenarios, optimizing the model architecture and training strategies, and integrating cuttingedge techniques to address challenges in encrypted traffic classification.

Differential privacy is a widely adopted privacypreserving technique for data publication. However, existing methods typically apply uniform noise to the entire dataset, neglecting the fact that the sensitivity levels of different attributes in various datasets can vary significantly. This uniform approach often leads to unreasonable privacy budget allocation and diminished data utility. To address this issue, this paper proposes a data publication method based on multilayer sensitivity analysis for personalized differential privacy(MLSAPDP). The proposed method first designs a sensitivity scoring strategy that provides finegrained, comprehensive evaluation from the perspectives of individual attributes, tuples, and their interrelationships. Then, privacy budgets are personalized according to sensitivity levels. In addition, data clustering is used to group similar data, reducing global sensitivity and minimizing noise injection. This not only enhances privacy protection but also ensures high data utility. Experimental results demonstrate that compared to traditional differential privacy methods, the proposed approach more effectively protects sensitive data, achieving an optimized balance between privacy protection strength and data utility.

The energy trusted data network primarily addresses challenges in achieving trusted interconnection, intercommunication, interdiscovery and interoperation of data, supported by digital object architecture (DOA) technology, and enables unified crossentity data access registration, directory interconnection services, and controllable analytical applications, flexibly meeting the development requirements of the energy industry, which demands high data security, organizational hierarchy, and multidomain entity segmentation. With reference to the traditional PESTEL (political, economic, social, technological, environmental, legal) environmental analysis model and the legaltechnologicaleconomiccommercial system model for data factor market development, this study proposes a systematic research framework for the energy trusted data network mechanism. Centered on data characteristics, the framework integrates dimensions of policy systems, industry layout, innovative technologies, and security compliance. Guided by the foundational principles of “costeffectiveness, equivalence of rights and responsibilities, collaborative integration, and longterm development”, it establishes a distributed overarching architecture. The mechanism is further constructed through the following aspects: collaboration mechanisms, technological mechanisms, incentive mechanisms, operational mechanisms, security mechanisms, and iterative mechanisms, to support energy industry advancement, flexible technological upgrades, and optimized evolution. By building this trusted network, more entities are encouraged to securely unify data access and leverage trusted service applications, transforming fragmented enterprise data advantages into industrywide collaborative strengths. This fosters deeper industry data utilization and advances artificial intelligence large language model development, providing critical support for the digital transformation and highquality development of the energy sector.

In response to the challenges faced by traditional cybersecurity detection and assessment technologies—such as large system scales, dynamic supply chain risks, and insufficient evaluation depth—this paper explores the application of AI technologie to advance this field. Methodologically, an endtoend implementation framework for largescale models is proposed, consisting of “data preparationdistillation and annotationcluster trainingquantitative deployment.” A localized compliance assessment model based on retrievalaugmented generation (RAG) technology is developed, and a multimodal model supporting joint textimage analysis is deployed. The large model significantly shortens the assessment cycle in scenarios such as provincial government clouds, improves the efficiency of compliance knowledge matching while reducing computational load by 70%, and markedly enhances the detection rate of inherent defects. The conclusion indicates that AI technology can effectively overcome the limitations of traditional assessment methods, promoting cybersecurity detection and assessment toward greater intelligence, adaptability, and comprehensiveness, thereby providing support for building resilient cybersecurity protection systems and fostering related ecosystem development.

With the increasing number of network attacks, network traffic anomaly detection is becoming more and more important for maintaining network security and stability. However, existing methods are often difficult to effectively capture both static statistical features and dynamic temporal features of network traffic during feature extraction, resulting in limited detection performance in complex and evolving network environments. To address these issues, this paper proposes a twochannel multiscale gated fusion anomaly detection model (MSAD) based on semisupervised learning. The model first extracts  static statistical features of the traffic, including the number of packets, total bytes, etc., through a multiscale convolutional neural network. Secondly, the temporal features of network traffic data are captured through a bidirectional GRU network and combined with a multihead attention mechanism. Finally, adaptive fusion of different modal features is performed through gated fusion mechanism. Meanwhile, for the problem of insufficient credibility of pseudolabel generation in semisupervised learning, a twostage adversarial pseudolabel generation strategy is proposed, which effectively improves the robustness of pseudolabels. The experimental results show that under the condition of limited labeled data, the model proposed in this paper achieves 99.63%, 99.54%, 99.9% and 99.72% of accuracy, precision, recall and F1 value on the CICIDS 2017 dataset, which is significantly better than traditional machine learning and deep learning methods.

In order to address the issues of coarse granularity in existing dark Web intelligence classification research and the predominance of Englishlanguage datasets, this paper proposes a finegrained analysis study focused on Chinese dark Web content. To overcome the scarcity of Chinese dark Web data and the misalignment of multimodal data, this study employs a large language model prompt rewriting strategy and a differentiated image enhancement strategy to achieve text and image data augmentation. By integrating product data from a certain platform on the Surface Web, a dataset comprising 14,052 product records was constructed. A feature selection optimization module was designed to establish an intertask coupling mechanism, and a Chinese dark Web product detection and classification model based on multimodal data augmentation was proposed. Experimental results demonstrate that the proposed model achieves macroF1 scores of 0.992 and 0.941 in dark Web product detection and classification tasks, respectively, representing an approximately 2% improvement over the best baseline model in  classification task and significantly outperforming existing singlemodal and multimodal methods. This approach effectively enhances the performance of finegrained classification tasks for Chinese dark Web intelligence, offering new insights and methodologies for dark Web intelligence analysis.