Deep neural networks face storage and computational bottlenecks when deployed on resource-constrained devices. Structured pruning techniques can effectively achieve model compression and acceleration by removing redundant weights，but the adversarial robustness of traditional pruning networks is insufficient，limiting their application in security-sensitive scenarios. To balance the needs for model lightweighting and robustness enhancement，an iterative optimization method combining adversarial training and structured pruning was proposed：during the adversarial training process，the pruning mask is optimized synchronously，and an adaptive training-pruning frequency adjustment mechanism based on the “exploration-exploitation”strategy was innovatively designed to realize the dynamic optimization of hyperparameters. Experimental results on the CIFAR-10 dataset and ResNet-18 model show that，under a sparsity of 0.7，the proposed method increases the model's robust accuracy by 10.32%; in extreme scenarios where sparsity exceeds 0.9，the normal accuracy and robust accuracy are improved by 4.76% and 15.52% respectively; compared with the fixed-frequency strategy，the adaptive mechanism further enhances the normal accuracy by 0.80%～3.59% and the robust accuracy by 1.30%～8.50%，significantly reducing the cost of manual hyperparameter tuning. This research provides an effective technical solution for the secure and efficient deployment of deep neural networks on mobile platform.

In modern military and civilian fields，UAV path planning in complex electromagnetic environments encounters severe challenges，such as dynamic distribution of threat areas，difficulty in balancing safety and efficiency with traditional algorithms，and insufficient accuracy in three-dimensional spatial modeling. To address these issues，an improved A^* path planning algorithm based on electromagnetic grid was proposed. Relying on the GeoSOT-3D framework，the electromagnetic grid conducts multi-dimensional and refined modeling of the electromagnetic environment，supports the dynamic fusion of multi-source information，and provides high-precision environmental representation for path planning. On the basis of the traditional A^* algorithm，this algorithm introduces a probability factor to reconstruct the loss function，quantifies the threat cost using radar detection probability，and combines 3D diagonal distance to optimize heuristic search，thereby achieving a dynamic balance among path length，safety，and computational efficiency. Experimental results show that compared with the traditional A^* algorithm，the proposed method reduces the number of threat area grids by 46.94% and improves computational efficiency by 36.97%; it also supports adaptation to different task requirements by adjusting the danger coefficient. This study provides a highly robust path planning solution for UAV combat missions in complex electromagnetic environments.

In large-scale cyberspace mapping，rapidly and accurately detecting node information and identifying the operational status of devices is one of the core research contents. Currently，the version iteration speed of cyberspace devices is accelerating，and a large number of new-type devices are constantly emerging. How to track and identify the device type of the measured node has become a new challenge that needs to be solved urgently. Aiming at the problem that current research relies too much on existing knowledge and cannot adapt to device upgrade changes，a node device type identification method based on large language model（LLM）and retrieval-augmented generation（RAG）technology was proposed. First，relevant data were collected from RFC documents and Internet device manufacturer websites，and a knowledge vector database was constructed based on the embedding model.Then，the detected node feature information was encoded，and relevant background knowledge was retrieved from the vector database. The retrieved knowledge and node feature information were jointly constructed into prompts for the LLM. The reasoning ability of the LLM was used to identify the device type of the probed node. Finally，the effectiveness and performance of the method were verified through ablation experiments and real-network tests.

The rapid development of visual generative artificial intelligence（AI）technology has strongly driven innovation in fields such as artistic creation and medical image generation. However，its highly realistic generation characteristics also pose security challenges，including the spread of disinformation and privacy violations—thus，there is an urgent need for efficient detection technologies. To address the current issues of generative image detection methods，such as insufficient generalization on unseen data distributions and inadequate utilization of the text semantic potential of vision-language models，this study proposed a generated image detection method based on conceptual prompt-tuning，leveraging contrastive language-image pre-training（CLIP）.This method extracts prominent concepts in a data-driven manner to explore the common distributional features between generated images and real images，and generates semantic prompt vectors to inject rich prior knowledge into the CLIP text encoder. By optimizing the prompt vectors through prompt-tuning and a prompt ensembling strategy，it balances computational efficiency and the retention of pre-trained knowledge while enhancing detection capabilities across different models and datasets. Experimental results show that the proposed method significantly and consistently improves the performance of generated image detection，with average accuracy and precision on unseen domains increased by 5.96% and 6.37%，respectively. Additionally，it exhibits good robustness against common post-processing. Ablation experiments further verify the effectiveness and advancement of the proposed method，demonstrating its potential and reliability in practical applications.

Web application firewall（WAF）is critical defensive mechanisms against persistent threats，yet its security assessment has long been challenging. Traditional manual testing methods are inefficient and resource-intensive，while existing reinforcement learning（RL）based methods suffer from two major limitations：first，attackers cannot perceive the opaque rule logic of WAF，leading to low efficiency in black-box testing; second，the Boolean feedback of WAF causes the problem of sparse/delayed rewards—sparse rewards tend to trap intelligent agents in blind exploration，and delayed rewards hinder the association between early actions and final outcomes，seriously impairing learning efficiency. To break through these bottlenecks，this study proposed“Ouroboros”—ablack-box WAF testing framework—for the first time.Its core lies in converting the extracted WAF rules into an interpretable recurrent neural network（RNN）to provide fine-grained confidence scores，and integrating these scores with outcome-level rewards to drive RL-based testing.Experiments show that this framework can achieve a maximum bypass success rate of 89.2% on feature-based WAF. This not only alleviates the sparse reward problem and provides an efficient black-box testing solution，but also offers important references for optimizing WAF rules.

With the rapid popularization of Internet of Things（IoT）technology in various fields，network device identification has become a key link in the network security protection system. Real-time detection of IoT devices accessing the network is crucial for network management，security protection，and performance optimization. Accurately understanding network dynamics and identifying these IoT devices is a necessary prerequisite for effectively defending against hacker attacks. Traditional machine learning-based identification methods not only suffer from low efficiency，complex feature selection，and poor environmental transferability，but their accuracy also fails to meet the needs of practical protection. To address this issue，an IoT device identification method based on packet-level traffic semantic feature-enhanced large language models（LLM）was proposed. First，complex and heterogeneous IoT traffic was converted into universal packet-level traffic semantic features. Then，these packet-level traffic semantic features were used to fine-tune the LLM，enabling the LLM to automatically learn the potential traffic features of IoT devices and make device classification and identification decisions，thereby realizing end-to-end and efficient IoT device identification. Experimental results on the public datasets Aalto，UNSW，and hybrid CIC IoT datasets（2022，2023）show that the proposed method can effectively identify IoT devices based on packet-level traffic semantic features，and its the average identification accuracy can reach 99.99%，99.42%，and 98.83% respectively.

In the context of image adversarial attacks，white-box attacks targeting the target model often yield the best performance. However，in practice，it is usually difficult to obtain the architecture of the target model，which makes improving the transferability of adversarial examples particularly crucial. To address this issue，a training method based on generative adversarial network（GAN）was proposed to generate adversarial examples with strong transferability.The study finds that images themselves possess model-agnostic vulnerabilities，and generative methods implement attacks precisely by exploiting this characteristic. Unlike traditional methods that perform fine-tuning within the neighborhood of the original image，this method generates images with maximum likelihood from the distribution of other categories. These images are visually close to real images but can effectively mislead classifiers. During the training process，the generator produces adversarial examples，while the discriminator judges the correctness of their labels. The two components optimize collaboratively，continuously enhancing the adversarial potency and authenticity of the examples. Experiments show that the attack success rate of generative adversarial examples on multiple models is significantly higher than that of traditional methods，with an average improvement of approximately 25%，demonstrating stronger cross-model generalization ability. This result indicates that generative adversarial attacks not only enhance the practicality of black-box attacks but also reveal the widespread vulnerabilities of deep models，providing directions for the design of subsequent defense mechanisms.