검색 상세

Federated Multi-domain Learning for Intelligent Applications : Federated Multi-domain Learning

지능형 응용을 위한 연합 다중 도메인 학습

초록/요약

With the rapid development of communication technology and IoT, Federated Learning (FL), as an emerging distributed machine learning paradigm, realizes col- laboration and knowledge sharing among large-scale geographically distributed data or models while protecting data privacy, and is widely used in many fields such as intelligent surveillance, automotive networking, and healthcare. However, FL still faces many challenges in edge intelligence scenarios. First, there is a wide variety of terminal devices, and the types and quality of data collected are uneven. Traditional distributed learning usually handles all data indiscriminately and ignores the domain variability among data, leading to heterogeneity of data, i.e., the Non-Independent Identical Distribution (Non-IID) problem. Second, due to factors such as device energy consumption, arithmetic limitations, and communication anomalies, nodes frequently join or exit during FL training or inference, forming a dynamic topological network, which in turn increases the heterogeneity and robustness requirements of the system. In addition, existing anomaly detection methods tend to assume consis- tent data distribution, ignoring the subtle differences in data distribution in different domains, especially in scenarios such as network intrusion detection, 6G network data analysis (NWDAF), and Internet of Vehicles (IoV), all of which pose challenges. Therefore, how to optimize the FL framework for heterogeneous data, dynamic net- work topology, and privacy preservation requirements, focusing on the performance of multi-domain data, becomes a key issue to improve its performance. To this end, we firstly proposes a Federated Multi-Domain Learning (FMD) framework, which improves the generalization ability and communication efficiency of FL in heterogeneous data environments by introducing key techniques such as multi-domain modeling, duality optimization, and dynamic cluster partitioning. For the multi-domain network intrusion detection problem, this paper designs a detection model integrating deep auto-encoder (AE) and generative adversarial network (GAN), and optimizes the data distribution of different domains by using Lagrangian duality modeling method, which enables the anomaly detection model to identify anoma- lous behaviors in various domains more accurately. In addition, in order to reduce the communication overhead, this paper proposes a novel parameter transmission scheme, which enables federated training to significantly reduce the communication data volume while ensuring the detection performance. Secondly, we investigates NWDAF in 6G networks and proposes a multi-task optimization architecture based on FL to address the model fairness and robustness of NWDAF in multi-tasking and Non-IID data environments. Aiming at the conflict relationship among different NWDAF client tasks, we also introduces the global alter- nating gradient projection (AGP) optimization method to keep the model’s balanced performance among multiple tasks and improve the generalization ability of network anomaly detection. Experimental results show that the proposed framework outper- forms traditional methods in detection on multiple NWDAF tasks and significantly reduces the communication cost. Thirdly, an FMD-IoV framework is proposed to address the heterogeneity and privacy protection of multi-view camera data for IoV in an intelligent transportation scenario. The framework groups data with similar viewpoints by clustering method and maps different types of image data to a unified feature space using multi-domain learning to reduce the impact of data heterogeneity on model training. Meanwhile, for the dynamic network topology changes in Telematics, we proposeed model de- normalization and similarity aggregation strategies to improve the adaptability and generalization ability of FMD-IoV. Experimental results show that the MSE of this method is significantly reduced on Synthia and CityScape datasets, verifying its ef- fectiveness and robustness in large-scale IoV environments.

more

목차

I. Introduction 1
1.1 Motivation and Objectives 1
1.2 Research Contributions 3
1.2.1 Communication-efficient Multi-domain Network Anomaly De-tection 4
1.2.2 Multi-task 6G NWDAF Network Anomaly Detection 5
1.2.3 Security and Robust Enhancement for FMD-IoV 5
1.2.4 Relationship Between Different Proposals and Thesis Focus 6
1.3 Organization of the Dissertation 7
II. Background 8
2.1 Federated Learning 8
2.1.1 Distributed Machine Learning 10
2.1.2 Parameter Server 11
2.1.3 Federated Edge Intelligent 12
2.1.4 Federated Tool 13
2.2 Multi-domain Machine Learning 14
2.3 Federated Multi-domain learning 15
2.3.1 Federated Multi-task learning 16
2.3.2 Federated Transfer Learning 17
2.4 3GPP 5G Analysis Framework 18
2.5 Internet of Vehicles 19
III. Federated Multi-domain Learning for Network Anomaly Detection 21
3.1 Introduction 21
3.2 Proposed Work 24
3.2.1 Multi-domain Machine Learning Model 24
3.2.2 FL Multi-domain Framework 26
3.2.3 Robust Optimization for FMD 28
3.3 Clustered Framework 30
3.3.1 CFMDML-i 32
3.3.2 Communication Optimization 34
3.4 Experiments 36
3.4.1 Datasets 36
3.4.2 Settings 37
3.4.3 Results 38
3.5 Summary 49
IV. Fair Federated Learning for Multi-task 6G NWDAF 51
4.1 Introduction 51
4.2 Proposed Approach 54
4.2.1 MTL fairness 54
4.2.2 Federated NWDAF Multi-task 58
4.2.3 AGP 62
4.2.4 Proof of the AGP Convergence 64
4.3 Experimental Setup and Performance Evaluation 68
4.3.1 Experimental Setup 68
4.3.2 Performance Evaluation 70
4.4 Summary 78
V. Federated Multi-domain Learning for IoV 80
5.1 Introduction 80
5.2 Proposed Framework 83
5.2.1 Model Designed 83
5.2.2 Model Convergence 87
5.3 Federated Learning for IoV 88
5.3.1 FMD-IoV 88
5.3.2 Model Weight Similarity Aggregation (FMD-MWSA) 90
5.3.3 Federated de regularization(FMD-de-reg) 94
5.4 Experimental Setup and Performance evaluation 96
5.4.1 Datasets 96
5.4.2 Baseline 97
5.5 Performance Evaluation 98
5.5.1 Performance of Federated Vehicle Networking 100
5.5.2 Ablation Study 103
5.6 Summary 104
VI. Conclusion and Future Work 109
6.1 Conclusion 109
6.2 Future Work 110
References 113

more
반출 Meta View 목록