INTRODUCTION: Distributed networks generate heterogeneous security telemetry while moving data across endpoints, users, services, and operational domains. SOCs need methods that protect data assets, preserve auditability, and avoid unsafe tool calls.OBJECTIVES: This paper proposes a data-model dual-driven method, in which incident and execution data constrain LLM-based reasoning while model outputs generate auditable process data, for privacy-aware data security decision analysis in multi-agent security operations.METHODS: The method combines LLM-based role agents, SOAR playbook orchestration, persistent message state, and a virtual security capability layer. Incidents are transformed into data-aware tasks, actions, commands, execution records, and summaries.RESULTS: On 83 labeled incident samples, tool-call evaluation achieved 0.9684 precision, 0.4742 recall, 0.6367 F1-score, and 76.45 s average handling time.CONCLUSION: The method supports auditable data security monitoring and controlled response, while complex multi-step planning remains the main improvement target.
INTRODUCTION: Dynamic Industrial Internet of Things (IIoT) delivery is characterized by stochastic order arrivals, heterogeneous UAV–UGV capabilities, and real-time demand updates, making conventional static routing methods insufficient for collaborative delivery planning. Existing UAV trajectory-planning and truck–drone routing studies usually focus on either single-platform trajectory control or offline vehicle routing, and they rarely integrate stochastic IIoT demand updating, UAV–UGV rendezvous coordination, operational constraints, and multi-objective decision-making into a unified reinforcement-learning framework.OBJECTIVES: To address this limitation, this study proposes a guidance-factor-enhanced MADDPG framework for collaborative multi-objective path planning of UAVs and UGVs in random-demand and dynamic delivery environments.METHODS: Order arrivals are modeled by a Poisson process, while order location, payload, priority, and time-window attributes are generated through multivariate random distributions and updated by IIoT terminals every 5 s. Payload, endurance, speed, rendezvous, and task-sequencing constraints are incorporated into a normalized multi-objective function considering delivery time, delivery cost, demand satisfaction, and UAV endurance loss. A distance–demand–priority guidance factor and a reconstructed reward function are further introduced to alleviate sparse-reward effects and guide agents toward high-value demand regions.RESULTS: Simulation and ablation results show that the reconstructed reward is essential for stable learning, while the guidance factor significantly accelerates convergence. Under 40 IIoT nodes, the proposed method converges after approximately 2,000 episodes, whereas the model without guidance information requires about 5,000 episodes. Comparative experiments further show that the proposed method improves system throughput and collaborative delivery efficiency while maintaining feasible UAV trajectories.CONCLUSION: The proposed framework provides an adaptive reinforcement-learning solution for UAV–UGV cooperative logistics in dynamic IIoT environments.
INTRODUCTION: In Industry 5.0 manufacturing systems, adaptive decision support is required to handle demand fluctuations, machine degradation, and operational feasibility constraints under human supervision. However, existing digital twin and data-driven approaches remain primarily focused on monitoring or forecasting, with limited integration into decision-oriented production control and feedback adaptation. OBJECTIVES: This study aims to develop a human-in-the-loop adaptive production decision-support framework that transforms external demand signals into feasible, risk-aware, and reviewable production actions by integrating demand-production state fusion, digital twin-based constraints, structured generative decision decoding, and feedback-driven updates. METHODS: A constrained decision-making system is constructed using Online Retail II and AI4I 2020 datasets. A unified state representation is learned by fusing demand indicators with manufacturing digital twin variables. A constraint-guided policy decoder is designed to generate discrete production actions under feasibility constraints rather than performing demand forecasting. Human feedback is encoded into subsequent decision states to enable iterative adaptation. The model is trained via a unified optimization objective balancing demand alignment, risk control, and explanation consistency. RESULTS: Experimental results on two datasets show that the proposed framework achieves improved demand-production alignment while consistently reducing manufacturing risk exposure and human override rates compared with forecasting-based and decision-oriented baselines. CONCLUSION: The proposed framework demonstrates that integrating digital twin constraints, constrained generative decision-making, and human feedback enables a shift from demand prediction to feasibility-aware production decision optimization, improving safety, interpretability, and human acceptability in Industry 5.0 manufacturing systems.
INTRODUCTION: With the rapid expansion of cross-language text processing applications, distributed systems must support large-scale multilingual workloads with high efficiency and responsiveness. Effective load balancing and real-time scheduling are therefore essential for maintaining system performance. However, many existing approaches rely on static or general-purpose scheduling strategies, which fail to capture language-dependent workload variations, resulting in limited scalability and inefficient resource utilization. OBJECTIVES: This study aims to improve load balancing efficiency and real-time scheduling performance in distributed systems for cross-language text processing. The focus is on addressing dynamic task variations, heterogeneous resource demands, and robustness challenges in multilingual, noisy, and fluctuating computing environments. METHODS: A cross-language-aware dynamic load balancing and scheduling optimization framework based on reinforcement learning is proposed. The framework incorporates language pair, task type, priority, input length, estimated resource demand, and real-time node status into an adaptive scheduling strategy. Denoising mechanisms and feature fusion modules are further introduced to enhance scheduling stability and decision robustness. RESULTS: Experimental results show that the proposed method outperforms existing approaches in response time, resource utilization, and task completion rate. The framework achieves a task completion rate of 96.5%. Under high-noise conditions, the reduction in task completion rate is approximately 50% lower than that of baseline methods, demonstrating improved robustness. CONCLUSION: The proposed framework provides an effective solution for dynamic scheduling and load balancing in cross-language text processing systems. By coupling multilingual task characteristics with distributed resource states, it offers scalable support for multilingual distributed computing and new insights into reinforcement learning-based scheduling optimization.
INTRODUCTION: In rail transit equipment manufacturing, which encompasses locomotive body welding, metro vehicle assembly, and high-speed rail carriage production, high-fidelity 3D point cloud data acquired by industrial sensors serves as the foundation for digital twin modeling, automated quality inspection, and robotic guidance. However, harsh production environments characterized by metallic dust, welding spatter, mechanical vibration, and frequent occlusions by fixtures and tooling inevitably introduce severe data corruption, including missing regions and non-uniform point density. Such degraded sensor data undermines the reliability of downstream manufacturing processes that depend on accurate 3D representations. OBJECTIVES: This paper presents an effective 3D sensor data recovery method that reconstructs complete geometric representations from corrupted partial scans, specifically targeting the data integrity challenges encountered in rail transit equipment manufacturing. The proposed approach aims to simultaneously restore global structural completeness and local geometric precision, thereby enabling resilient digital twin systems that maintain operational continuity despite sensor-induced data loss. METHODS:We propose an Attention and Residual Aware Network (ARaNet) featuring a Multi-Scale Channel-Aware Convolution encoder and a Hierarchical Residual-Aware Decoder. The encoder performs Farthest Point Sampling at multiple resolutions (2048, 1024, 512 points) and applies a channel attention mechanism to dynamically weight feature channels, emphasizing geometrically salient structures such as sharp edges, curved surfaces, and mechanical joints that are critical in rail transit component geometries, including bogie frames, car body shells, and coupler assemblies. The decoder progressively generates point clouds from coarse resolution (64 points) to medium resolution (128 points) to high resolution (2048 points), incorporating a residual refinement module that predicts coordinate offsets to rectify local geometric errors. This capability is particularly valuable for precision metrology in component manufacturing. RESULTS: Experimental evaluation on the ShapeNet and ModelNet40 datasets demonstrates that ARaNet achieves substantial improvements over benchmark models PCN and PF-Net. Quantitative assessment shows average reductions of 12.9% and 23.8% in the Gt-to-Pre and Pre-to-Gt metrics, respectively. The generated point clouds exhibit more uniform distributions and superior detail restoration, particularly for complex mechanical geometries with curved surfaces and structural joints characteristic of rail transit equipment components. Conclusion:The effectiveness of integrating channel attention with residual refinement mechanisms for industrial 3D sensor data recovery is validated. ARaNet provides a robust data recovery foundation for resilient digital twin systems in rail transit equipment manufacturing, enabling sustained operational capability and rapid geometric reconstruction even when sensor data is compromised by production environment disruptions. However, it should be noted that ARaNet is currently validated exclusively on synthetic training data, and its performance under extremely high missing-region ratios or in real industrial deployment scenarios without domain adaptation warrants further investigation.
The agricultural supply chain is facing practical challenges such as data privacy breaches and insufficient cross domain data collaboration control in distributed scenarios. This article proposes a collaborative management mechanism that integrates distributed blockchain, time series analysis, and privacy computing to study data security in distributed agricultural supply chains. This article first elaborates on the theoretical basis of relevant technologies, with a focus on how blockchain achieves privacy isolation for cross node data transmission through asymmetric encryption technology, uses smart contracts to achieve dynamic permission control and operation tracing of distributed nodes, and adapts to the dynamic monitoring needs of distributed data streams with the real-time advantage of time series analysis; On this basis, a management mechanism covering overall architecture, node collaboration, anomaly monitoring, and cross domain data flow control was designed, and an algorithm model including data preprocessing, blockchain node feature extraction, privacy protection data processing, time series anomaly detection and analysis was constructed. Through experimental verification, accuracy, recall rate, RMSE, MAE and other indicators were evaluated using one-year operational data from a certain agricultural supply chain scenario as a sample. The results showed that the accuracy of the experimental group was 92%, the recall rate was 88%, the RMSE was 12.5, and the MAE was 9.8, all of which were better than the control group. Research has shown that this collaborative mechanism and algorithm model can effectively enhance the distributed data security protection, cross domain data flow control, and anomaly recognition capabilities of agricultural supply chains, solve the data security and privacy protection problems of agricultural supply chains in distributed scenarios, provide new methods for the safe and stable operation of agricultural supply chains, and have important engineering practical value and promotion significance.
One of the principal requirements for Intelligent Manufacturing Systems that are working in complicated, nonlinear, and uncertain environments is the robust real-time monitoring and an adaptive control system. In this paper, a deep reinforcement learning-based adaptive control framework is proposed that combines multi-sensor fusion, residual-based anomaly detection, and robust state estimation. The framework is tested on a multi-sensor manufacturing dataset that consists of a total of 10,000 time steps of four different heterogeneous sensor streams collected under both normal and anomalous operating conditions. The sensor data is initially preprocessed by noise filtering, normalization, synchronization, and feature extraction, after which an Unscented Kalman Filter (UKF) is used for sensor fusion and state estimation. Residual analysis along with Support Vector Machine (SVM) allows for real-time anomaly detection and classification, while Multi-Model Adaptive Estimation (MMAE) technique improves the robustness of the state estimation. A Proximal Policy Optimization (PPO) agent uses the updated system state and the anomaly information for adaptive control. The results of the experiments reveal perfect anomaly classification with an ROC-AUC of 1.000, highly stable state estimation in all state dimensions, and continuous control performance increase with the average reward of 1.79, thereby validating the proposed framework's suitability for anomaly-aware adaptive control in intelligent manufacturing systems.
INTRODUCTION: Patent texts may present challenges in accurately identifying potential patent infringement risks due to subtle differences in technical features. OBJECTIVES: Deep learning effectively models the complex semantic structures and local feature correlations within patent texts by integrating deep semantic representations and deep matching mechanisms, thereby enhancing sensitivity to minute technical variations and improving decision support for infringement risk assessment. OBJECTIVES: this study explores deep learning-based optimization methods for patent infringement risk text comparison and retrieval algorithms. A macro-level patent text comparison and retrieval layer is constructed by integrating Word2vec and LDA topic models. This layer models patent text semantics, generates word-level semantic vectors for patents, and enables rapid comparison and retrieval of candidate patent sets related to the target patent at the technical feature semantic level. A COV-BiGRU twin neural network is introduced as the fine-grained comparison and retrieval layer. This layer optimizes the macro-level retrieval algorithm by performing granular semantic matching and Manhattan distance calculations between candidate patent texts and the target patent text. Based on these results, patent infringement risk retrieval are retrieved. RESULTS: Results demonstrate that this method reduces the candidate patent text set size to 0.457% of the full database during macro-level retrieval. CONCLUSION: In the fine-grained retrieval optimization phase, it achieves 100% recall of known high-risk patents while effectively excluding non-infringing patents that share similar technical themes but differ in specific technical features. This validates the preliminary effectiveness of the optimized dual-layer retrieval algorithm for texts with fine-grained differences on the tested patent corpus. This validates the preliminary effectiveness of the optimized dual-layer retrieval algorithm for texts with fine-grained differences on the tested patent corpus.
INTRODUCTION: Distributed digital art platforms face ownership ambiguity, privacy leakage, unauthorized access, and audit difficulties in cross-domain data flows. OBJECTIVES: To enhance trust, privacy, controllability, and auditability through a dedicated data security governance mechanism.. METHODS: A closed-loop governance mechanism jointly models art data, AIGC models, copyrights, and transactions, integrating five coordinated modules: damage-tolerant ownership verification, privacy-preserving cross-chain auditing, dynamic access control, secure delivery with rollback, and efficient traceability. RESULTS:Ownership confirmation reaches 97.9%-98.7% (mild perturbations) and 85.9% (combined); authorization accuracy: 95.6%; illegal access interception: 94.9%; normal transaction audit pass rate: 99.3%; anomaly detection: up to 100%; on-chain storage compression: 99.99% (50 MB); traceability success: 94.8%. Access latency remains below 20 ms with 180+ transactions/s under concurrency. CONCLUSION: The mechanism effectively secures cross-domain data flows with high trustworthiness and low overhead, suitable for distributed digital art ecosystems.
INTRODUCTION: With the growth of Internet of Things (IoT) and edge computing, distributed systems are increasingly deployed across fields such as sports health and intelligent transportation. However, motion data, often containing sensitive personal information, poses significant privacy risks when handled in these systems. OBJECTIVES: This paper aims to propose a novel security-aware scheduling method that integrates motion data privacy protection in distributed systems. The goal is to balance system scheduling efficiency with robust privacy safeguards. METHODS: We introduce a framework that combines encryption technologies, privacy protocols, and dynamic scheduling algorithms. By embedding privacy protection constraints into the scheduling process, this method optimizes data transmission and storage during task execution. RESULTS: Experimental results demonstrate that the proposed approach effectively reduces privacy leakage risks by over 80% compared to classical greedy algorithms. While the mandatory cryptographic mechanisms introduce a marginal latency overhead, the system maintains highly competitive scheduling efficiency. When compared with state-of-the-art techniques such as pure DRL, the proposed system achieves a 71% reduction in privacy leakage probability, successfully balancing robust security with dynamic adaptability. CONCLUSION: This research presents an innovative solution for motion data privacy protection in distributed systems, offering significant improvements in both privacy and scheduling performance. The method's applicability extends to fields like IoT, smart health, and intelligent transportation, marking a crucial step toward more secure and efficient distributed systems.
INTRODUCTION: Nowadays, object detection and recognition play an important role in various applications such as surveillance, autonomous driving, robotics, and medical imaging. However, none of the traditional works focuses on analyzing the explicit relationships, logical dependencies, and semantic conflicts in text-rich or complex scenes, affecting object recognition accuracy. OBJECTIVES: A Natural Language Processing (NLP)-based robust object detection and recognition framework is developed through multimodal relation graph construction using Minimum Persistence Spanning Tree (MPST) and Deep Swim Wishart Distribution Convolutional Neural Network (DSWDCNN). METHODS: Initially, image with their corresponding captions is collected. Then, the image preprocessing and text preprocessing are done independently. From the preprocessed image, the object is detected using You Aspect-ratio Adaptive Anchors Only Look Once version-8 (YAAAOLOv8), followed by visualization. Meanwhile, from the preprocessed text, entity relations are identified. The multimodal relation graph is constructed using MPST. Further, the features from preprocessed text, relation graphs, detected objects, and visualized-images are extracted. Next, the multimodal analysis is carried out. RESULTS: In the meantime, the word embedding is performed on the preprocessed texts. Finally, the object recognition is carried out using DSWDCNN. CONCLUSION: The proposed framework achieves an object recognition accuracy of 98.8569%, demonstrating its effectiveness under weakly supervised conditions.
This paper proposes a universal post quantum privacy protection edge identity authentication framework to address the challenges faced by edge identity authentication in distributed cross domain networks, such as quantum attack threats, cross domain data privacy breaches, and difficulties in coordinating anonymity protection and compliance supervision. The framework adopts an optimized lattice based linkable ring signature protocol to meet the lightweight operation requirements of edge nodes and prevent the risk of leakage in identity data interaction; Design traceability constraints and controllable cross domain traceability mechanisms based on the linkability feature of signatures, balancing user privacy and regulatory requirements. Prove that the scheme possesses unforgeability, strong anonymity, and quantum resistance under the random oracle model. After optimizing the algorithm and interaction logic, the authentication efficiency is improved by 8% to 15% compared to similar solutions, and it is adapted to the low computing power and low latency characteristics of edge nodes. Combining zero knowledge proof to build a lightweight data collection mechanism and achieve privacy protection throughout the entire data process. This article uses the integrated aviation tourism system as a typical application case to verify that the proposed framework can be widely applied to various distributed cross domain networks and identity authentication systems.
Continuous operation in tobacco warehouses places high requirements on circular-rail rail guided vehicle (RGV) dispatching, because task concentration, track interference, and equipment abnormalities may quickly cause vehicle waiting and material-flow delay. To improve the response capability of this system, this study develops a hybrid scheduling method based on an improved genetic algorithm and simulated annealing. According to the operating characteristics of the circular rail, a scheduling model is formulated to minimize the overall task completion time while considering rail operation rules and composite task requirements. In the solution process, the simulated annealing temperature is used to adjust parent selection and mutation behavior, so that the algorithm can maintain search diversity in the early stage and improve convergence in the later stage. Case results show that the proposed method reduces the average maximum completion time from 150.2 s to 143.4 s and improves the best result from 145.7 s to 139.6 s compared with the conventional genetic algorithm. These results indicate that the proposed method can improve RGV response efficiency and provide decision support for stable warehouse logistics under disturbances such as task fluctuation and path conflict.
INTRODUCTION: Urban renewal digital twin systems must support high-fidelity rendering and millisecond-level interactive control across geographically distributed, heterogeneous edge computing infrastructures. Conventional scheduling approaches, often assuming hardware homogeneity and static task graphs, struggle with dynamic service-chain reconfigurations driven by urban spatial entropy fluctuations, leading to resource misallocation and service instability in large-scale distributed environments. OBJECTIVES: This paper aims to develop a real-time, scalable scheduling framework for heterogeneous distributed edge systems that jointly accounts for hardware diversity, evolving directed acyclic graph (DAG)-based workloads, and decentralized decision-making while preserving data locality and model privacy. METHODS: We propose a hierarchical scheduling architecture integrating physical and logical coordination: (1) a resource affinity mask encodes hardware constraints as a prior to prune infeasible placements; (2) a spatio-temporal graph neural network captures critical path dynamics in non-stationary task DAGs; and (3) a federated policy distillation mechanism enables knowledge transfer across structurally diverse cloud-edge-end agents without sharing raw models or data. RESULTS: Experiments on the Alibaba Cluster Trace and Shanghai Telecom datasets show that the proposed method reduces average latency to 43.8 ms, achieving a 25.6% reduction compared with CO-MARL, sustains a 94.7% task completion rate under long-tail traffic surges, and achieves an edge inference latency of 2.1 ms. CONCLUSION: The “physical priors plus topology awareness” paradigm demonstrates that heterogeneous-aware, distributed coordination is essential for real-time digital twin services at scale.
INTRODUCTION: The reliability of decision-making in adaptive artificial intelligence (AI) systems is limited by uncertain factors such as noise in multi-source sensing data and conflicts in decision-making objectives. Furthermore, in distributed multi-node collaborative environments, challenges such as cross-node data leakage and edge-node privacy risks further exacerbate decision uncertainty. OBJECTIVES: To address this issue, an uncertainty-aware decision-making method for adaptive AI systems based on robust Bayesian networks is proposed, with a specific focus on distributed data security and privacy protection. METHODS: A robust Bayesian network-based model is proposed. Multi-source sensing data are fused using adaptive weighted minimum mean square error optimization. Secure distributed aggregation is achieved through encrypted aggregation and differential privacy mechanisms. Federated transfer learning is introduced to prevent raw data sharing, and minimax estimation is used to correct missing-data bias. RESULTS: Experimental results on distributed inspection robot clusters show that the proposed method effectively achieves uncertainty-aware decision-making under dynamic and obstacle-affected scenarios while maintaining privacy protection and CONCLUSION: The proposed framework improves decision robustness and privacy preservation in distributed adaptive AI systems by integrating Bayesian inference, federated learning, and secure data fusion strategies.
This paper presents a novel Numerical Index Modulation-assisted Differential Chaos Shift Keying (NIM-DCSK) system, architected on a multi-carrier framework where numerous information-bearing subcarriers are supported by a single common reference subcarrier. The proposed system introduces an additional dimension for data transmission by embedding information bits into the dynamic allocation of signal energy across subcarriers. This mechanism not only boosts the transmission rate but also significantly enhances physical layer security. Specifically, the NIM-DCSK scheme obfuscates its transmission characteristics through pseudo-random energy allocation patterns, where deliberately varying power levels create a complex and non-stationary signal signature. This fluctuating energy distribution effectively masks the underlying data structure, making it extremely difficult for unauthorized eavesdroppers to perform feature analysis or achieve successful decoding. Compared to conventional DCSK systems, comprehensive analyses demonstrate that the NIM-DCSK system achieves substantial gains in both energy efficiency (EE) and spectral efficiency (SE), alongside improved bit error rate (BER) performance under various channel conditions. These advantages are realized without a commensurate increase in structural complexity. The inherent security features, including inherent resistance to passive interception and a reduced vulnerability to statistical analysis, are key benefits. Consequently, this work facilitates the integration of index modulation with chaos-based communications, ensuring high-rate transmissions coupled with robust security. The proposed system is, therefore, highly suitable for applications demanding reliable and secure wireless links, such as secure sensor networks, tactical military communications, and confidential IoT deployments.
We, the Publisher, have retracted the following article: Fan et al. (2025). The Quality Evaluation of Innovation and Entrepreneurship Education in Colleges and Universities in the Context of Big Data. EAI Endorsed Scal Inf Syst. https://doi.org/10.4108/eetsis.7015 The article has been retracted due to authorship dispute. We informed the authors about this decision. The retracted article will remain, and it has been watermarked as “RETRACTED”.
The rapid evolution of the Internet of Things (IoT) demands robust architectural models capable of integrating emerging technologies and managing growing system complexity. This paper presents a structured review of IoT Reference Architectures (RAs), analysing their conceptual foundations, structural organization, and technological readiness in the context of Next-Generation IoT (NGIoT) technologies. The study examines RAs based on their support for disruptive technologies, including Edge/Fog/Cloud Computing, 5G, Artificial Intelligence, Digital Twins, Augmented Reality, Blockchain, and the Tactile Internet. Special emphasis is placed on the ASSIST-IoT RA, evaluated as a modular, cloud-native blueprint aligned with modern software design principles. Reflecting decentralization, scalability, interoperability, and resilience, ASSIST-IoT emerges as a production-ready framework for building intelligent and adaptive IoT systems. The findings synthesize current RAs trends and limitations, offering a forward-looking perspective that informs the development of NGIoT systems driven by data-driven and human-centric innovation.
INTRODUCTION: As digital transformation progresses, power communication networks become exposed to increasingly complex security threats. Moreover, traditional routing algorithms cannot perceive or respond to dynamic security threats, making reliable data transmission difficult to achieve under network attack conditions. OBJECTIVES: Therefore, this study proposes a risk-aware secure routing (RASR) mechanism for power communication networks based on graph neural networks (GNNs). METHODS: The mechanism first introduces an autoencoder-graph neural network (AGNN) architecture for determining critical nodes, thus establishing a scoring prediction model customized to the structure of power communication backbone networks. It then integrates the essential scores of the nodes, historical failure rates, and traffic load factors to devise a path node risk quantification model. Finally, it incorporates risk quantification results into routing policies to enable proactive routing avoidance and thus bypass high-risk nodes. RESULTS: Experimental results show that the RASR algorithm effectively meets the differentiated service requirements of heterogeneous traffic-including delay-sensitive, bandwidth-sensitive, and reliability-sensitive traffic. Compared with traditional routing algorithms, it demonstrates greater stability and fault tolerance under high-risk attack scenarios while eliminating the need for redundant backup strategies, thereby markedly reducing resource overhead. CONCLUSION: Therefore, the proposed mechanism offers important theoretical support and a novel technical approach for establishing secure, reliable power communication networks.
Data heterogeneity, the complexity of privacy budget allocation, and the imbalance between privacy and performance lead to a limited scope of privacy protection constraints and fail to ensure data integrity. Therefore, an edge computing communication privacy protection method based on federated learning algorithm is proposed. Participants use federated learning to locally train the sensing data to obtain a local model, avoiding the interaction of raw data with edge computing nodes and the perception platform. The parameter values of the trained model are perturbed by noise using the adaptive differential privacy technology and uploaded to the edge computing node. The edge computing node performs edge aggregation on the noisy model parameters and uploads them to the perception platform to complete the global aggregation operation, realizing edge computing communication privacy protection. A performance loss constraint mechanism suitable for federated learning is proposed and designed, and the performance loss of the adaptive differential privacy federated model is reduced by optimizing the constraint scope of the loss function, improving the privacy protection effect. Experiments show that this method can effectively add noise to the local model parameters and achieve privacy protection for edge computing communication; the performance loss of this method’s privacy protection is small, about 0.4; when transmitting different types of data in edge computing communication, the data integrity after privacy protection processing by this method is above 0.98, with excellent privacy protection effect.
