We present , a multi-agent rule generation system designed for industry-scale payment fraud detection. combines classical pattern mining with constrained, evidence-driven LLM reasoning to automatically generate, validate, and refresh interpretable symbolic rules within predefined risk score bands. The closed-loop system turns statistical patterns into structured hypotheses and controlled generalizations, deterministically evaluated on historical data, while preventing the LLM from accessing raw transactions or computing performance metrics. Using real transactional data sampled over multiple months as testing data, we observe significant improvements in false-positive and fraud-capture metrics within targeted score regions.
Resource scheduling in cloud computing environments is a highly complex task. On the one hand, the complexity stems from the diverse optimization challenges posed by various cloud infrastructures; on the other hand, user requests in these environments are highly dynamic. Ineffective scheduling strategies can lead to increased task processing times, which in turn may result in a higher number of Service Level Agreement (SLA) violations. Such performance degradation directly undermines cloud security by creating resource starvation for security agents (e.g., intrusion detection systems) and prolonging task exposure windows, thereby increasing the attack surface. Although heuristic algorithms have been widely used, they often exhibit significant limitations when dealing with large-scale systems, thereby diminishing their practical effectiveness. To address these challenges, this paper proposes a scheduling method based on the Double Deep Q-Network (DDQN) framework. The proposed method is designed for a discrete action space and incorporates a dynamic reward function capable of adaptively reflecting the quality of current scheduling decisions. While directly optimizing for performance metrics, this function inherently promotes a more stable and predictable resource environment, which is a fundamental prerequisite for robust cloud security. Experiments are conducted in a simulation environment built using SimPy, evaluating the proposed approach under low, medium, and high workload scenarios. The method is benchmarked against FirstMax, Random, A2C, PPO, and DQN algorithms. Performance is assessed based on metrics including makespan, average task completion time, average weighted turnaround time, and number of SLA violations. Experimental results demonstrate that the DDQN-based scheduling algorithm consistently outperforms all baseline methods across all workload levels. Compared to the second-best performing method, FirstMax, the proposed approach reduced makespan by an average of 1.15
Standard digital signatures are publicly verifiable, but certain applications require special privacy-preserving measures. Therefore, there is a need to prevent signatures from being maliciously disseminated by unspecified users. This paper proposes a strong designated verifier proxy signature scheme based on lattice (L-SDVPS), which prevents the theft of privacy in signatures by non-designated verifiers. The L-SDVPS scheme is based on the small integer solution (SIS) lattice problem, which is resistant to quantum computation attacks. The proposed scheme is anonymous and unforgeable under the random oracle model (ROM). Additionally, the scheme has a shorter signature size and a lower storage consumption advantage compared to related schemes.
Temporal knowledge graphs (TKGs) play an increasingly prominent role in dynamic scenarios such as social network evolution, financial forecasting, and smart city operations, where capturing evolving relationships is critical. Temporal knowledge graph representation learning holds significant practical value by modeling dynamic semantic evolution. However, significant challenges, which is reflected in effectively encoding temporal patterns and integrating structural dependencies across evolving graph states, exists in current approaches. To meet these challenges, we propose a novel framework named as temporal knowledge graph representation and learning with evolutional representation and attention mechanism that synergizes two complementary innovations within an encoder-decoder architecture. The first innovation involves evolutionary representation learning, which dynamically captures the evolution of entities and relations by integrating structural dependencies within temporal snapshots through RGCN operations. The second innovation focuses on modeling sequential patterns across adjacent states. The time-aware attention mechanism adaptively weights historical contexts through attention layers, amplifying salient temporal correlations for query-specific prediction. Through these components, the proposed TKGRAM can improve key issues in existing methods. Extensive experiments on ICEWS benchmarks demonstrate TKGRAM’s consistent superiority, outperforming state-of-the-art baselines, and further, ablation studies validate both innovations as critical drivers of performance gains.
Deep learning-based object detection models have been widely applied in autonomous driving, industrial quality inspection, and other fields, with security issues triggered by adversarial perturbations receiving significant attention. However, current adversarial attack and defense methods for object detection models predominantly focus on single attack or defense strategies. This limitation hinders their ability to counter diverse unknown attacks, resulting in poor attack adaptability and insufficient defensive generalization capabilities of the models. To address these challenges, this paper proposes an integrated adversarial attack and defense method for object detection models. By integrating multiple mainstream attack methods to generate diversified adversarial examples, and combined with adversarial training for defense. Experiments conducted on both general object detection and specific object detection tasks demonstrate that the integrated attacks significantly diminish model accuracy, while the integrated defense mechanism promptly restores model performance. This approach effectively enhances model security in complex adversarial scenarios.
