The field of cybersecurity operation faces challenges such as fragmentation of knowledge, low response efficiency, and professional data sensitivity. In order to better meet the above challenges, local fine-tuned vertical domain Large Language Models (LLMs) for cybersecurity—LEAD-Cyber are proposed based on open-source LLMs and full-cycle training datasets. A multi-step generation method is used to build a professional knowledge dataset in the field of cyber security. It meets the needs of three training stages of LLMs: pre-training, instruction fine-tuning and reasoning fine-tuning. The DeepSeek and QWen open-source LLMs were also optimized in full cycles using full-parameter fine-tuning methods and low-rank adaptation (LoRA). Based on subjective and objective indicators, we use evaluation indicators such as ROUGE, BLEU and WinRate analysis to evaluate the performance of LLM on different benchmarks to verify the effectiveness of LLM in handling different tasks. Experiments show that the LLM after fine-tuning is significantly better than the baseline model. The research verifies the advantages of the full-cycle fine-tuning strategy in optimizing the field of knowledge expression and maintaining general capabilities, providing an efficient and reliable solution for intelligent and cybersecurity operation and maintenance.

The consensus on the state consistency of Unmanned Aerial Vehicle (UAV) swarms is a crucial guarantee for the collaborative operations. Existing state consistency consensus methods are mostly applied in scenarios such as distributed databases and blockchain, characterized by stable network topology and simple task requirements. However, the high dynamics of UAV swarms lead to issues such as fragile network connections and diverse coexisting states, making it difficult for existing state consistency consensus methods to be applicable. To address these issues, this paper proposes a state consistency consensus architecture for UAV swarms, providing theoretical support for ensuring the collaborative capabilities of UAV swarms. Firstly, it analyzes the problems existing in the state consistency of UAV swarms in terms of models, protocols, and networks, summarizing existing research. Secondly, based on the Basically Available-Soft state-Eventually consistent (BASE) and Atomicity-Consistency-Isolation-Durability (ACID) theories, it establishes a relationship between soft state and weak consistency, and designs an asynchronous consensus protocol for multi-cluster soft forks under weak connections. Finally, simulation experiment results conducted in a Flying Ad-hoc Network (FANET) show that the proposed architecture can meet the state consistency consensus requirements of high-dynamic UAV swarms.

With the extensive application of UAVs in social production, their security issues have become increasingly prominent, particularly in the realm of communication and network security. Due to the deficiency of security mechanisms in design, the airborne CAN bus network is prone to be exploited by malicious devices, and the communication data may be tampered with or monitored, thereby posing severe security threats. The fundamental cause of this problem lies in that the initial design of the CAN bus prioritizes communication efficiency and low resource consumption, but neglects the security requirements, and is incapable of coping with the current complex security situation and diversified attack means. Additionally, the constrained resource environment of the UAV platform makes it challenging to directly apply traditional authentication and encryption technologies to the CAN bus network. To address this issue, an intrusion detection approach for the CAN bus network of UAVs based on an enhanced generative adversarial network (GAN) is proposed. This approach utilizes the game mechanism between the generator and the discriminator to generate pseudo-samples to enhance the training effect of the discriminator and improve the detection performance. Based on this approach, an experimental framework is established, and its validity and feasibility are verified in the resource-constrained environment. The experimental results show that the enhanced GAN model improves accuracy, recall, and F1 score by an average of about 5.56%, 3.93%, and 4.34% compared to the other three advanced deep learning models, respectively. This demonstrates its efficiency and reliability in drone CAN bus intrusion detection, providing important technical support and reference value for drone system security.