To address the issues of hallucinations in large language models for cyber threat attribution and the insufficient robustness of traditional graph-constrained reasoning against fragmented threat intelligence data, research on credible cyber attack traceability methods was conducted. The framework integrating threat intelligence knowledge graph and graph-constrained reasoning (TIKGCR) was proposed, which constructs an attack pattern-oriented knowledge graph Trie (KG-Trie) index and designs an adaptive soft-constrained decoding and multi-path evidence aggregation mechanism. Experimental results show that the framework achieves attribution accuracies of 94.6% and 92.8% on the CTI-Trace and TRAIL datasets respectively, with a hallucination rate of only 1.5%. It maintains an accuracy of 85.5% under a 50% knowledge graph missing rate and inference time cut by roughly 23. This framework achieves a balance between structural credibility and semantic robustness, providing effective technical support for cyber threat attribution in high-risk scenarios.
In new-type power systems, deep learning-based load forecasting models have been widely applied to resource scheduling and market response, making backdoor attack a more insidious potential threat. However, in these strongly coupled, multivariate forecasting scenarios, backdoor triggering not only needs to achieve targeted manipulation of the forecast output but also needs to consider key samples, variable selection, and cross-variable dependencies. To address these issues, a backdoor injection method , KoopTrigger, based on Koopman spectral sensitivity identification and graph structure perturbation modeling was proposed. Firstly, this method selects highly vulnerable input windows based on Koopman operator sensitivity analysis to generate efficient poisoning samples; then, it constructs a gated sparse graph convolution generator to inject low-rank perturbations into the variable dimensions, achieving structured backdoor triggering across windows; finally, a two-stage adversarial training mechanism is used to stably induce the model to learn the attack target. Experiments were conducted on four multivariate datasets: ETTh1, Ireland, Panama, and Milan. The results show that KoopTrigger can achieve efficient and covert prediction manipulation in multiple prediction models, significantly outperforming the ablation version. It also shows better attack performance and generalization performance compared to the baseline method.
With the rapid advancement of Industry 4.0, Intelligent manufacturing has become increasingly important. Traditional manual inspection still suffers from low efficiency, high cost, and limited stability, while supervised learning methods rely heavily on accurately labeled anomalous samples, which are scarce and diverse in real industrial scenarios. In contrast, unsupervised industrial image anomaly detection learns the distribution and intrinsic patterns of normal samples only, and has therefore attracted growing attention from both academia and industry.First, a systematic review of unsupervised industrial image anomaly detection under the one-class learning setting was provided. Existing methods were categorized into two main paradigms according to the source of anomaly scores: feature embedding and image reconstruction, and the development of representative approaches was summarized.Second, based on this taxonomy, the evolution and research trends of these two paradigms were analyzed. The common characteristics and differences among existing methods in terms of feature representation, anomaly scoring mechanisms, and localization strategies were summarized to provide a structured understanding of current research.Then, commonly used benchmark datasets and evaluation metrics in this field were reviewed.Finally, the main challenges in current research were discussed from the perspective of practical industrial applications, and potential future research directions were outlined to facilitate further development and deployment of unsupervised industrial image anomaly detection methods.
The differential analysis of block ciphers based on deep learning has become a hot research topic in the field of cryptography. To investigate the impact of different data formats on the performance of neural network differential distinguisher, by combining the characteristics of single-differential and multi-differential distinguisher data formats, and taking advantage of the differential propagation uncertainty of the S-box component in the CRAFT, a “single-differential + S-box intermediate state” distinguisher data format was designed and proposed. Through residual neural network (ResNet) models, three types of data formats corresponding to the 5-7 rounds of neural network differential distinguisher of the CRAFT were obtained. For the 5-round neural network differential distinguisher, the accuracy of all three discriminators was above 97%. For the 6-round neural network differential distinguisher, the accuracy of the “single-differential + S-box intermediate state” distinguisher was 96.14%, significantly better than the 61.39% of the single-differential distinguisher and the 72.78% of the multi-differential distinguisher. For the 7-round neural network differential distinguisher, the accuracy of the three distinguishers was slightly above 50%. Among them, the training data complexity of the model corresponding to the single-differential distinguisher was 221.07, while the training data complexity of the multi-differential and “single-differential + S-box intermediate state” distinguisher was 221.66. Further, by using a 6-round “single difference + S-box intermediate state” distinguisher to conduct partial key recovery attacks on the 7-round CRAFT, when the number of plaintext-ciphertext pairs selected was 10 000, the success rate can reach 93% in 100 attack experiments, the time complexity of a complete cryptanalysis attack was 222.01.
Fully homomorphic encryption does not support nonlinear homomorphic computation, making it difficult to directly implement homomorphic ReLU evaluation in privacy-preserving machine learning. Existing schemes that employ high-degree polynomials to approximate the ReLU activation function involve large-bitwidth modular multiplication and high-degree polynomial multiplication operations, resulting in limited system throughput and low resource utilization.To address this, a CPU-FPGA heterogeneous collaborative architecture was proposed for homomorphic ReLU evaluation, with the CPU executing high-degree polynomial approximation and the FPGA accelerating homomorphic polynomial evaluation. Specifically, a segmented lookup table modular reduction algorithm was first designed. By means of a two-level decomposition mapping mechanism, large-bitwidth modular multiplication was converted into parallel high/low-bit lookup table operations and linear combinations. As a result, the modular multiplication delay was reduced by 97% compared to the existing state-of-the-art scheme. Secondly, a dynamic resource-reusing CG-NTT architecture was proposed to optimize polynomial multiplication efficiency. Relying on a fixed memory access pattern, a dynamic table reuse strategy was implemented. Compared with traditional NTT architectures, 15%～37% of the computation units were reduced, and 45%～50.8% of the operation latency was decreased. Finally, a CPU-FPGA heterogeneous acceleration framework was constructed. Through a phase-tuning interface and a resource coordination mechanism, a speedup of 21.4 times was achieved for homomorphic ReLU evaluation. The feasibility of the proposed scheme in accelerating ciphertext inference was verified by experiments, although the most time-consuming bootstrapping operation still had to be executed on the CPU side. Therefore, future research will be focused on the hardware acceleration of the bootstrapping operation, in order to further improve the practical performance of the ciphertext inference system.
Mobile App have become an essential part of people's daily lives, but they also brought about security risks to data privacy. Some Apps share user data with third parties in the background or obtain user data from third parties to deliver personalized recommendations. Due to the stealthiness of this data sharing channel, it is difficult to be effectively tracked under the current regulatory system. In response to this issue, an intelligent mining solution was designed and implemented to automatically detect potential data sharing behaviors between Apps. The detection framework was constructed based on automated testing, which injects user preference data with keywords by simulating user behavior, and then collects screenshots of personalized recommendation interfaces. Information extraction and correlation analysis technologies based on AI models were introduced, using optical character recognition models, multimodal visual large models, and large language models to complete the intelligent extraction and understanding of text information and image content in App screenshots, and to make judgments on the correlation between keywords and image content descriptions. The injection of user preference data was performed on 30 Apps, resulting in a dataset of 8 700 screenshots, and yielding 1 466 pieces of evidence of data sharing. Among the 30 Apps, 23 were found to provide data to third parties, and 29 were found to receive data from third parties and engage in personalized recommendations. A network structure of App data sharing behavior was constructed, providing a new solution for the automated detection of App data sharing behavior.
Federated learning suffers from fairness imbalance due to gradient conflicts in non-independent and indentically distributed (Non-IID) data scenarios, with model parameter transmission risking privacy leakage. Existing fair federated learning algorithms mitigate gradient conflicts effectively but lack targeted optimization for local-global model consistency under complex heterogeneous data, leading to deviated local training from global optima, impaired fairness and limited overall performance. To address this, the FedALF algorithm and LDP-FedALF algorithm were proposed: it alleviates gradient conflicts via fairness-driven objectives and hierarchical optimization; adopts ADMM for optimization, designing dual variables and quadratic penalty terms in clients' local objective functions to enforce local-global model consistency constraints, further enhancing fairness and performance; and integrates adaptive local differential privacy into model transmission for privacy protection. Through a large number of experiments, FedALF achieved an average improvement of 10.82% in test accuracy and 19.80% in fairness performance compared to existing fair federated learning algorithms. LDP-FedALF achieved an average improvement of 7.53% in test accuracy and 15.36% in fairness performance compared to existing fair federated learning algorithms.
Existing post-quantum cryptographic algorithms have been confronted with high computational complexity and large storage overhead, while current blockchain systems have suffered from significant consensus throughput pressure in multi-node environments. To address these issues, a lightweight post-quantum distributed consensus scheme for blockchain systems was proposed in this work. The scheme was constructed by improving the FORS structure in the SPHINCS+ post-quantum signature algorithm, through which the signature size and storage burden were reduced, and by designing an improved WOTS+ structure to decrease the computational complexity of the algorithm. Furthermore, a grouped PBFT consensus mechanism was proposed, and a primary node election mechanism was designed by integrating a credit-based scoring strategy with a weighted random forest model, thereby reducing communication overhead and improving network reliability and throughput. Experimental results demonstrated that, compared with the traditional SPHINCS+ post-quantum scheme and improved PBFT algorithms, the proposed scheme achieved improvements of 25% and 28% in average signature verification rate and consensus throughput, respectively, while the signature storage pressure and consensus latency were reduced by approximately 31% and 25%. These results indicate that the proposed scheme effectively enhances signature and verification efficiency as well as overall system throughput of blockchain systems, while reducing consensus latency.
A study was conducted on intrusion path detection based on an improved hidden Markov model (HMM) to address the challenges of strong concealment of multi-step intrusion paths in the Internet of Things, scarce annotated data, and high proportion of encrypted traffic. A quadruple intrusion graph model was constructed that integrated network topology, configuration, and vulnerability information, state transition probability was quantified using logistic functions, a dynamic probability matrix was formed, and a maximum likelihood probability reallocation strategy was designed based on semantic similarity for zero probability defects in traditional HMM. Introducing state transition correction coefficients and adaptive dynamic thresholds, a multi-step iterative inference framework was constructed to achieve intelligent screening of alarm sequences and robust recognition of intrusion intentions. The experimental results show that the method has an accuracy of over 94% in path detection in complex attack scenarios, and has good robustness and scalability.
To address the efficient processing requirements of network streaming data in high-performance intelligent computing systems, a spatiotemporally parallel reconfigurable computing array architecture was proposed and applied to the design of intelligent chips. In this architecture, multi-context processing units and storage-enhanced micro-processing units were integrated, and dynamic context switching, near-memory computing, and pipeline parallelism were supported. The memory wall was broken through the deep coupling of computation and storage, and the resource utilization was improved by means of time-division multiplexing. Compared with traditional solutions, the utilization rate of processing units was increased by more than 30%, the latency was reduced by more than 25%, and the energy efficiency was significantly enhanced. In this paper, the parallel mechanism of the proposed architecture was elaborated in detail, the problems existing in traditional architectures, the mainstream technical approaches, and the innovations of this scheme were analyzed, and this key technology was further applied to the design of reconfigurable smart NIC chips. A new domestic solution for network transmission in scenarios such as intelligent computing centers and AI training and inference was thereby provided.
The service-based architecture (SBA) represents a pivotal innovation of the 5G core networks. It decouples traditionally tightly coupled network elements into loosely coupled, reusable, and independently deployable network functions (NF), which can be flexibly invoked via standardized service-based interfaces (SBI). This paradigm enables cloud-native deployment, modular design, elastic scaling, and open interoperability of mobile networks. Nevertheless, the openness and flexibility introduced by SBA also expand the threat surface for network security and user privacy. Recently, research efforts and standardization activities concerning security and privacy preservation for SBA mobile networks have been vigorously advanced. However, there still lacks a systematic survey that comprehensively sorts out the academic advances and research trajectories in this field. To fill this research gap, this paper conducts a systematic survey covering security, privacy, trustworthiness, and resilience issues in fully SBA networks. Firstly, we thoroughly analyzed the security and privacy risks faced by three main components of the mobile networks, i.e., the core network, the access network, and the user terminals, and summarized the key security and privacy requirements of the networks. Then, we proposed the four-dimensional trustworthiness evaluation criteria system (4DTECS), specifically designed for SBA network. By adopting 4DTECS, we performed a comprehensive analysis and investigation on existing academic research and standardization efforts. Furthermore, we identify several critical open challenges and outline some promising future research directions, aiming to provide valuable theoretical insights and technical references for academia and industry, and facilitate the sound and secure evolution of next-generation mobile communication networks.
Graph neural networks can effectively learn and detect maliciously implanted hardware trojans (HT) in gate-level netlists. However, existing methods were usually trained on isomorphic and same-process circuits, which resulted in the generalization ability of the models being limited by the process and structural characteristics of the training set, making it difficult to cope with the graph structure distribution shift caused by the differences between the target circuit and the training set. To address this issue, a gate-level HT detection method based on graph transformer was proposed. Firstly, the gate-level netlist was converted into a general graph structure, and a multi-dimensional feature extraction module was designed to capture the intrinsic attributes and topological information of nodes. Secondly, a three-layer stacked graph Transformer was adopted for feature encoding to capture the global dependencies between nodes, and multi-level features were fused to enhance the expression ability. Finally, a dual-task classifier with a gradient reversal layer was introduced to suppress the domain shift caused by circuit heterogeneity. Experiments at the Trust-Hub public library showed that the proposed method achieved a recall rate of 95.2% on the synopsys 90 nm general library SAED, and its F1-score was improved compared with existing mainstream methods. On the large-scale dataset based on TRIT, it achieved an average recall rate of 87.9% and an average F1-score of 85.0%. This method effectively improves the robustness of HT detection models in cross-process or heterogeneous netlists.
With the increasing demand for fair trading in the energy sector, the use of blockchain technology as a new model for energy trading has begun to come into the public eye. However, there are problems such as the lack of a third party to endorse, relatively low efficiency, and unreasonable distribution. The innovative integration of the dynamic proof of stake (PoS) mechanism and entropy regulation strategy in the energy trading sharding system was proposed to solve problems such as trust deficiency, low transaction efficiency and uneven energy distribution existing in traditional energy trading. In the optimized triangular model, verifiable random functions were adopted to select validators, and the staking weights were dynamically adjusted in combination with the real-time status of nodes and transaction activity. Quantify the distribution of equity using Shannon entropy and set a threshold to trigger redistribution. The Pareto frontier solution set for the three objectives of throughput, security and decentralization was solved through the NSGA. A mechanism was adopted to dynamically adjust the equity weight based on the real-time status of nodes and the activity level of energy transactions, effectively enhancing the efficiency and fairness of transaction verification. The regulation was introduced for quantification to reduce the uncertainty and chaos of the energy trading system, ensuring the rational allocation and efficient utilization of energy resources. The experimental results show that optimized system throughput and attack cost have been significantly improved. Meanwhile, the degree of decentralization has risen to 89%, and the overall performance has increased by 12.8 times. This scheme has advantages such as good transaction efficiency, security and energy conservation and consumption reduction, which is conducive to reducing energy transaction costs, enhancing the transparency and stability of the energy transaction market.
Federated learning faces dual challenges of data heterogeneity and privacy leakage in distributed privacy-preserving scenarios.Personalized federated learning mitigates data distribution discrepancies by constraining local model update directions,while differential privacy provides rigorous privacy guarantees.However,when solving these two problems at the same time,the existing methods fail to fully consider the coupling effect of noise on the federated learning and training process in the data Non-IID scenario,resulting in insufficient privacy-utility trade-offs.Experiments show that differential privacy noise significantly exacerbates the dimension collapse problem in-duced by data heterogeneity.This problem leads to the decay of effective degrees of freedom in the feature space and impairs discriminative feature representation,which is a key factor in model performance degradation.To ad-dress this,a differential privacy personalized federated learning framework incorporating feature decorrelation and residual constraints(DP-FedDRC)was proposed.This framework alleviates the dimension collapse problem caused by both noise and data heterogeneity by suppressing redundant feature correlations and preserving critical discriminative information.In terms of privacy protection,DP-FedDRC implements gradient clipping and Gaussian noise injection using the DP-SGD mechanism during local training,and rigorously accounts for the global privacy budget based on the Rényi differential privacy framework.Experiments on the MNIST,Fashion-MNIST,and CIFAR-10 datasets demonstrated that under equally strict privacy constraints,DP-FedDRC achieved higher model accuracy compared to baseline algorithms,with only an 8.3%～9.9%increase in computational overhead.
Non-orthogonal multiple access (NOMA) has been regarded as a key enabling technology for sixth-generation (6G) communications, since resource utilization was significantly enhanced by allowing multiple users to share the same time-frequency and spatial resources. However, in power-domain NOMA systems, user privacy information may be eavesdropped due to the application of successive interference cancellation, which introduces potential security vulnerabilities. To address this issue, various physical layer security schemes have been proposed in the literature, while their performance was highly dependent on sufficiently accurate channel state information (CSI). When channel estimation accuracy was insufficient, the achievable security performance often failed to meet expectations. To overcome this limitation, a physical layer security scheme based on channel estimation error compensation was proposed under imperfect CSI conditions. First, channel estimation errors were compensated through an additional low-complexity estimation process. Then, the transceiver was redesigned based on the more accurate CSI obtained after error compensation. Simulation results demonstrated that the proposed scheme significantly improved the secrecy rate and transmission reliability under different levels of channel estimation errors and signal-to-noise ratio. This scheme now provides a feasible solution to the physical layer security problem in scenarios where channel estimation errors exist.
With the continuous evolution of phishing attack methods, traditional detection approaches face severe challenges in terms of feature extraction and model generalization capability. To address these issues, a phishing detection model named DPFEAT (phishing detection model with dual-path feature extraction and adaptive fusion), which is based on dual-path feature extraction and adaptive fusion, was proposed. Firstly, a dual-path feature extraction mechanism for temporal and spatial features was constructed, and a feature interaction layer was introduced to enhance the overall feature representation capability. Secondly, an adaptive fusion module was designed to achieve dynamic fusion of multi-dimensional features, thereby effectively improving the model's ability to capture complex attack characteristics. Regarding the training strategy, a warmup-cosine decay and cyclical learning rate scheduling mechanism was adopted, which significantly improved the stability of the training process. Simultaneously, a combined loss function was constructed by integrating Focal Loss and Dice Loss, and a dynamic weight balancing mechanism was introduced, which effectively alleviated the issues of class imbalance and the difficulty in identifying boundary samples. Furthermore, an interpretability analysis framework was designed to evaluate the contribution of each module to the model's performance. Experimental results demonstrated that the proposed DPFEAT model achieved an accuracy of 95.41% and an receiver operating characteristicarea under the curve (AUC-ROC) score of 98.83%, which is significantly superior to existing methods.
Software vulnerabilities have long been a key threat to the field of network and information security, and vulnerability detection is the primary technical means to mitigate such threats. Vulnerability detection based on code similarity offers the advantages of scalability and low cost, particularly when combined with artificial intelligence technologies such as graph neural networks and recurrent neural networks, further enhancing its accuracy and applicability, making it a research hotspot in this field. This paper systematically reviews the progress of software vulnerability detection based on code similarity in the past seven years: from the perspective of datasets, it analyzes the categories, granularity, scale, and balance of vulnerability sample datasets; from the perspective of code preprocessing, it summarizes two types of methods: code cleaning and standardization, and disassembly and intermediate language representation; from the perspective of code representation, it categorizes six schemes of detection mechanisms, advantages and disadvantages, and applicable scenarios based on code snippets, hashes, trees, graphs, metrics, and hybrid representations, and compares their detection performance and overhead under unified evaluation metrics. Furthermore, it distills the categories and characteristics of the machine learning techniques employed. Finally, it summarizes the challenges faced by this research and looks forward to future research directions.
The onion routing (Tor) anonymous network, due to its multi-layer encryption and dynamic relay mechanisms, posed significant challenges for traffic classification. Illegal applications or websites traffic hidden within Tor anonymous network traffic could not be effectively detected by traditional traffic identification techniques. To address this issue a semi-supervised learning-based method for encrypted traffic classification in Tor networks was proposed. A unified "pretraining + fine-tuning" framework was utilized to fully leverage unlabeled data. In the pretraining phase, a dynamic masking mechanism was introduced to gradually adjust the masking ratio, enhancing the model’s ability to learn latent features. In the fine-tuning phase, a channel and spatial attention (CSA) module was integrated to improve the model’s sensitivity to critical local regions. The overall approach was built on a Transformer-based architecture, combining self-supervised feature extraction with enhanced discriminative capability. Experimental results on the ISCXTor2016 dataset and a self-collected dark web traffic dataset showed that the proposed method achieved over 98% across all performance metrics in classifying eight typical traffic categories, demonstrating excellent generalization ability and practical value.
Software-defined edge computing networks enable efficient, reliable and secure network communication. The topological architecture of software-defined edge computing networks requires flexibility and scalability to adapt to the technological environments including Internet of things, artificial intelligence and big data. Edge computing devices are interconnected through high-speed links, core network and converging nodes. In software-defined networks, the OpenFlow discovery protocol (OFDP) utilizes link layer discovery protocol (LLDP) packets for sending and receiving related messages and generating network topology. The edge node update rate is typically high under different scales and requirements while the core backbone convergence switches do not require repeated updates. Existing topology discovery schemes for software-defined edge computing networks consequently exhibit low efficiency in redundant link identification and topology reconstruction, ultimately causing controller overload and affecting edge computing task offloading mechanisms. To resolve these issues, a lightweight topology discovery method for software-defined edge computing networks was proposed. For redundant link discovery challenges, dynamic packet binding, verification and real-time forwarding were implemented through multi-level flow table rules and optimized LLDP frame protocols. For excessive message transmission during topology discovery in SDN controllers, an edge node and link discovery algorithm based on minimum vertex cover concept was developed. Additionally, an MVC-based dynamic single/double-directional edge link resolution and merging algorithm was introduced to alleviate message overload in SDN controllers. Experimental results demonstrated that compared with existing network topology discovery algorithms, the proposed method achieved lightweight network topology discovery while maintaining capabilities for resolving link capacities and inter-node connections during reconstruction. This approach effectively reduces data packet volume within the network and decreases controller workload.
Provenance graphs were widely used for detecting and investigating sophisticated network attacks.How-ever,their high storage and analysis overhead made graph compression indispensable.Most existing compression methods were found to disrupt critical causal dependencies among nodes,thereby compromising the effectiveness of attack investigations.In particular,current compression algorithms that preserved investigative capabilities struggled to handle redundancies arising from cyclic interactions between nodes.To address this challenge,a novel graph compression algorithm,CPC,was proposed.By thoroughly analyzing causal relationships and recording es-sential node dependency information,redundant transitive dependency edges were precisely identified and re-moved,significantly improving edge compression efficiency while preserving the integrity of attack investigation results.Two variants of the algorithm were developed:CPC-A,which retains full dependencies for offline prov-enance graph construction,and CPC-P,which preserves partial dependencies for online scenarios.Experimental evaluations on public datasets showed that,compared with current state-of-the-art compression algorithms,the pro-posed approaches achieved substantially higher compression performance,with overall compression ratios im-proved by 9.3%～86.4%.
