To improve the intelligence level of UAV security detection and address the privacy leakage and offline failure problems of cloud-based large models, this paper designed and implemented a lightweight UAV-side security multimodal large model, UAV-SMLLM. First, a multimodal security detection dataset containing 20000 dialogue instances and covering 10 attack types was constructed. Second, a layer pruning strategy based on block-level angular distance redundancy analysis was proposed, which compressed the LLaVA-1.5-7B-HF decoder from 32 layers to 24 layers. Finally, LoRA fine-tuning and AWQ 4-bit quantization were adopted, reducing the model’s peak memory footprint to 3.75 GB and achieving an inference speed of 5.08 tokens/s. Experimental results show that the model achieves detection accuracies of 90.75% and 84.87% for spoofing attacks and brute-force attacks, respectively; the intent recognition F1 score increased from 0.475 to 0.7945; and its general visual capability is on par with that of the baseline large model.

To address the privacy leakage, copyright disputes, and data transparency issues brought by large language models, this article systematically reviewed membership inference attacks and defense techniques. Taking the large model lifecycle and system components as two main lines: First, it explained the formal definition and threat model of membership inference, and elaborated on the theoretical basis of its evolution from overfitting to local and conditional memorization, clarifying the study’s incremental contribution compared with existing general surveys on membership inference. Then, it organized attacks in scenarios such as pretraining, fine-tuning, alignment, black-box application programming interfaces (API), multimodal, and retrieval-augmented generation (RAG), with emphasis on supplementing LiRA, MoPe, ReCaLL, and multiple types of RAG attacks. Furthermore, it quantitatively compared representative results of benchmarks such as WikiMIA and MIMIR, pointing out that the distribution bias caused by temporal splitting could significantly overestimate attack capability. Finally, it constructed a defense-in-depth framework from five levels: data governance, regularization and differential privacy, knowledge distillation, decoding filtering, and machine unlearning. The analysis shows that attack signals have shifted from global statistics to per-sample calibration, local memorization peaks, conditional changes, and system knowledge coverage, while strict in-distribution and low false-positive-rate evaluation remains a core issue constraining practical auditing.

Backdoor poisoning during large language model fine-tuning threatens model security, yet attacks, defenses, and evaluation practices remain fragmented. This review examined instruction tuning, domain adaptation, and parameter-efficient fine-tuning through a literature review, a taxonomic analysis, and a cross-study comparison. It classified attacks by poisoning timing, trigger pattern, and attack objective; It organized defenses into data processing, training-time constraints, backdoor diagnosis, and model remediation; And aligned key metrics such as attack success rate, clean utility, false trigger rate, stealthiness, and defense cost. The analysis showed that current studies focused mainly on static poisoning before training and fixed lexical triggers. Semantic, composite, and multi-turn triggers still lacked stable definitions and reliable detection methods. Most defenses depended on specific trigger assumptions, clean data, or model access, limiting transfer across models and tasks. The review concluded that comparable evaluation requires explicit threat models, data provenance, judging rules, and access conditions. Research on continual fine-tuning and black-box settings remains necessary to improve the comparability and credibility of security evaluations.

The surging scale and complexity of modern software systems posed severe challenges to traditional static analysis. To address this, this paper systematically reviewed the evolutionary trajectory, core challenges, and state-of-the-art methods in repository-level static vulnerability detection. First, it traced the generational technological evolution from rule-driven traditional static analysis and data-driven machine learning/deep learning to semantics-driven large language model(LLM) and agent, and analyzed the critical role of intermediate representations in cross-file modeling. Then, for repository-level scenarios, it summarized six core challenges, including excessive context scale, fragmented cross-file dependencies, and difficulty in controlling false positives. It systematically evaluated the capability boundaries of technical approaches such as the deep integration of static analysis tools with LLM and collaborative agent-based detection frameworks. Finally, on this basis, it deeply dissected the trustworthiness bottleneck caused by the conflict between probabilistic reasoning and formal constraints, identified several structural limitations in existing datasets, proposed an approach to reconstructing evaluation metrics for engineering effectiveness, and outlined future research directions.

To address the limited capability of traditional malware detection methods against evasion techniques such as packing and obfuscation, this paper proposed a malware detection method based on hardware performance counter. The method employed a Transformer encoder to perform global modeling on hardware performance counter time-series data, extracting deep temporal features that capture the dynamic evolution patterns of microarchitectural behaviors. An XGBoost classifier then performed the final classification decision based on these features, achieving a complementary integration of temporal inductive bias and structured inductive bias. Experiments were conducted on hardware performance counter time-series data collected from 516 malicious samples and 500 benign samples under short, medium, and long observation windows respectively to evaluate the proposed method. The results demonstrate that the framework achieves detection accuracy exceeding 95% across all three window configurations, and the performance coefficient of variation is lower than that of the comparison method, exhibiting favorable detection precision and cross-window stability. The above complementary mechanism effectively enhances the robustness of the classification boundary, offering a detection method that balances representational capacity and generalization performance for malware detection based on microarchitectural behavior analysis.

In recent years, the security of robust aggregation in blockchain-based federated learning become an important research topic. Existing defense methods mainly rely on detection in a single parameter space, prior trusted data, or fixed filtering strategies, and still suffer from limited detection dimensions and insufficient adaptability when facing stealthy attacks such as collusive poisoning and backdoor injection. To address these issues, this paper proposed MFDD, a robust aggregation method for blockchain-based federated learning based on multi-dimensional features and dynamic defense. MFDD identified potential collusive attacks through similarities among client updates and evaluated client anomalies using a probe dataset in terms of prediction deviation, task loss, parameter deviation, and directional consistency. To address directional anomalies in collusive reverse updates, a directional-risk conditional trust mechanism based on a robust center was further constructed. A coordinate-wise median was used to form a robust reference center, and an unsupervised median absolute deviation (MAD) criterion was employed to identify significant reverse updates. When valid directional-risk evidence was formed, anomalous clients were excluded and the weights of the remaining clients were renormalized. On this basis, client filtering and weighted aggregation were performed through adaptive fusion and dynamic defense, while key detection and aggregation information was recorded on the blockchain ledger to enable trusted recording and traceable auditing. Experimental results show that under the MNIST IID setting with 35% collusive reverse updates, MFDD achieves an accuracy of 0.9704± 0.0025, whereas the ablated variant without the directional-risk conditional trust mechanism achieves an accuracy of only 0.1763. On-chain auditing experiments based on Hyperledger Fabric 2.5.14 achieve 318 matches between on-chain and off-chain records and validated the auditing mechanism through duplicate submissions, overwrite attempts, and tampering with core audit fields. The results demonstrated that MFDD can improve robust aggregation under complex attack scenarios while providing a trusted and auditable recording mechanism for key detection and aggregation information.

To address prominent problems in modern enterprise data security governance, such as unclear asset inventories, inefficient cross-departmental collaboration, and delayed policy enforcement, this paper proposed a collaborative governance framework for data security risks, named CGF-DSR. Taking data assets as the core governance object, the framework constructed a unified data asset and risk view model and, based on a policy-as-code engine, enabled automated execution and dynamic evaluation of security and compliance policies. In addition, it designed a cross-functional collaborative workflow mechanism to conduct closed-loop management of risks throughout the data lifecycle. Practical application shows that the CGF-DSR framework can effectively break down collaboration barriers among data, business, security, and compliance departments, significantly improve the efficiency of data risk identification, handling, and auditing, and enhance enterprises’ data security governance resilience and agility in multi-cloud environments.

With the development of audio-visual generation techniques, deepfake videos increasingly exhibit localized, short-duration, and multimodal manipulation patterns. Short-term forgery cues may be weakened by numerous normal frames during global temporal modeling. To address this issue, this paper proposed a window-guided audio-visual deepfake detection method for local temporal anomalies. Based on temporally aligned audio and visual features, our model constructed local dynamic information and cross-modal difference information, and organized the sequence into a frame-window-global hierarchy. The model preserved short-term anomalous responses within local windows, and enhances frame-level representations through cross-window contextual modeling and global semantic feedback, thereby balancing local anomaly perception and video-level authenticity prediction. Experimental results show that the proposed method achieves intra-dataset AUC scores of 99.54%, 99.90%, and 100% on FakeAVCeleb, DFDC, and LAV-DF, respectively. Cross-dataset and cross-manipulation experiments further evaluate the model under changes in data distribution and manipulation type. Ablation studies further verify the contributions of local-window modeling, anomaly-peak extraction, and global semantic feedback. Additional boundary-cropping and backend-efficiency experiments show that the model maintains stable performance under boundary perturbations, while window-level contextual modeling effectively reduces the computational and memory costs of backend long-sequence modeling. These results demonstrate that window-guided hierarchical aggregation preserves local short-term forgery cues while supporting video-level discrimination and efficient backend modeling of long sequences.

Indirect call is widely used in system software to implement dynamic program behavior, but it hinders the construction of accurate control-flow graphs, which underpins many static analyses and system-hardening techniques. Modern compilers can’t identify indirect-call targets by default, while most existing methods rely only on type information and produce many false positives. This paper proposed a program semantics-driven indirect call target recognition method called HRefIC. Due to the fact that effective targets were easier to reach the call point, the statistical results of 1000 randomly selected effective targets in Linux kernel, Bash, and MySQL showed that 57.4% of the call distances are 0, and 81.9% of the call distances do not exceed 2. There was usually a correlation between the caller, the called, and different objectives of the same indirect call, with 78.95% of valid objectives being similar to the caller’s function name. HRefIC integrated data flow, path, semantic, and text similarity information based on this, further refining the coarse-grained target obtained from type analysis. At the same time, on-demand design was adopted to control overhead, and multi-layer rules were used to reduce false alarm rates. The experimental results show that compared with traditional type analysis based on function signatures, HRefIC reduces the average number of targets by 98.6%, 94.4%, and 95.0% on Linux kernel, Bash, and MySQL, respectively, and further reduces by 38% to 66% compared to MLTA. In addition, HRefIC can also serve semantic error detection, using this method to discover 8 new errors in the Linux kernel.

The model context protocol provides a standardized interaction paradigm for tool invocation by LLM-based agents and enables compatibility with third-party servers. However, this open access model makes it difficult to constrain the evolving behavior of agents during multi-step tool invocation, increasing the uncertainty of runtime permission boundaries and consequently introducing risks such as unauthorized access and permission drift. This paper proposed MCP-DAC, a dynamic access control method that integrated fine-grained attribute-based authorization with event-flow graph constraints. During the trial phase, MCP-DAC automatically collected benign tool invocation logs, generated policy-based access control rules, and constructed an event transition graph for each session; at runtime, it performed attribute validation and path validation for each MCP request. Prototype experimental results with MCP Servers in multiple scenarios show that MCP-DAC effectively blocks various anomalous requests while causing limited interference with normal tool invocation workflows.

wireless sensor networks (WSN) have a wide range of applications and are vulnerable to malicious attacks. Focused protection of some nodes from network viruses can improve the overall network destruction resistance and fault tolerance. With the help of Holling-II class predation model, this paper quantitatively studied the dynamic behavior of a class of WSN worm propagation models with focused protection objects and designed targeted virus suppression strategies. Firstly, the local and global asymptotic stability conditions of the model were given. Secondly, the sufficient conditions for the model to generate Hopf bifurcation were investigated, and the direction of the bifurcation as well as the stability of the periodic solution were discussed. Finally, the parameter thresholds for the transformation of the model equilibrium point are derived, and the virus control scheme was designed accordingly. Theoretical analysis and numerical experiments show that the model can provide reference for the prevention and control of network worms.

To address the misuse of domain name system over HTTPS (DoH) for establishing covert channels, as well as the limitations of existing detection methods in cross-layer correlation and robustness to changes in class distribution, a multi-level behavioral feature fusion detection method, MBF-DoH, was proposed. The method integrated DNS query, TLS handshake, and traffic temporal features, and combined a bidirectional long short-term memory network with a DoH semantic relation graph to perform joint spatiotemporal detection. It further incorporated class-sensitive loss, relation preservation, and prototype constraints to adapt to emerging attacks. On the public dataset and the self-collected dataset, F1-score is 98.59% and 86.60%, and balanced accuracy is 97.80% and 96.92%, respectively. It can be seen that after the additional update, the F1-score of the new attack reached 95.27%, and the F1-score of the previously learned attack only decreased by 0.46 percentage points, which has stable detection power and the ability to continue to respond.

Ethereum faces increasingly complex fraud threats, while existing phishing detection methods mostly relied on static analysis of raw transaction networks, overlooked fine-grained dynamic transaction sequence patterns of accounts, and failed to fully exploit the synergistic benefits of structural and sequential features. Therefore, this paper proposed a TransG4Eth framework, which combined a transaction-level Transformer with a graph neural network to capture both transaction context information and account interaction relationships. The proposed method first constructed a Transformer-based transaction sequence encoder, which modeled the complete transaction history of accounts as fine-grained behavioral features. Second, it designed a contextual address recovery pre-training task, which guided the model to learn implicit interaction patterns among accounts in transaction sequences. Then, it constructed a global account interaction graph to explicitly characterize structural topology. Finally, it proposed an implicit-explicit dual learning framework, which achieved two-stage fusion at both the embedding level and the output level. Experimental results show that TransG4Eth significantly outperforms existing methods on real-world datasets, obtains an F1 Score of 85.99%.