- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Toward AI-native 6G networking : algorithm design for...
Open Collections
UBC Theses and Dissertations
UBC Theses and Dissertations
Toward AI-native 6G networking : algorithm design for efficient edge learning, caching, and inference Ma, Manyou
Abstract
The International Telecommunication Union (ITU) vision for the sixth generation (6G) networks includes the integration of artificial intelligence (AI) to provide intelligent and real-time services across diverse applications. To meet the demands of low-latency, privacy-sensitive, and computationally intensive tasks, deploying AI at the network edge is crucial. This approach enables data to be processed closer to its source, leveraging the computational resources at the edge. It reduces the latency and enhances both security and efficiency. The four major components of edge AI, namely, edge caching, edge training, edge inference, and edge offloading, are essential for managing AI workloads effectively at the edge. This thesis examines how the 6G wireless networks can support AI applications, including the key aspects of model training, distribution, and inference. First, we address the challenge to accelerate federated learning (FL) training in the scenario where the client datasets are non independent and identically distributed (non-IID). We show that when gradient diversity-based FL training is employed, the convergence of FL training depends on the age of information (AoI) of the model updates received from different clients. We propose a Lagrangian index-based solution and a two-stage client scheduling algorithm that jointly consider the gradient diversity, AoI, and channel condition of the clients. The proposed algorithm demonstrates substantial improvement in convergence speed of two FL tasks, achieving up to 71% reduction in average uplink transmission duration. Next, we explore dynamic content caching in heterogeneous networks (HetNets) to minimize the AoI of dynamic contents being cached. We formulate the problem as a constrained Markov decision process (CMDP) and identify a threshold structure of the optimal solution. We design a deep deterministic policy gradient (DDPG) algorithm that exploits the threshold structure of the solution. Our proposed method achieves up to 30% reduction in AoI compared to a heuristic strategy. Finally, we propose the Disentangled REgion-of-interest Attention Map (DREAM) framework for privacy-aware region-of-interest (RoI) segmentation in edge inference systems. By incorporating attention mechanism and adversarial training, the DREAM architecture protects sensitive user information in segmentation tasks, reducing the eavesdropper accuracy by up to 38% without significantly compromising the segmentation performance.
Item Metadata
| Title |
Toward AI-native 6G networking : algorithm design for efficient edge learning, caching, and inference
|
| Creator | |
| Supervisor | |
| Publisher |
University of British Columbia
|
| Date Issued |
2024
|
| Description |
The International Telecommunication Union (ITU) vision for the sixth generation (6G) networks includes the integration of artificial intelligence (AI) to provide intelligent and real-time services across diverse applications. To meet the demands of low-latency, privacy-sensitive, and computationally intensive tasks, deploying AI at the network edge is crucial. This approach enables data to be processed closer to its source, leveraging the computational resources at the edge. It reduces the latency and enhances both security and efficiency. The four major components of edge AI, namely, edge caching, edge training, edge inference, and edge offloading, are essential for managing AI workloads effectively at the edge.
This thesis examines how the 6G wireless networks can support AI applications, including the key aspects of model training, distribution, and inference. First, we address the challenge to accelerate federated learning (FL) training in the scenario where the client datasets are non independent and identically distributed (non-IID). We show that when gradient diversity-based FL training is employed, the convergence of FL training depends on the age of information (AoI) of the model updates received from different clients. We propose a Lagrangian index-based solution and a two-stage client scheduling algorithm that jointly consider the gradient diversity, AoI, and channel condition of the clients. The proposed algorithm demonstrates substantial improvement in convergence speed of two FL tasks, achieving up to 71% reduction in average uplink transmission duration. Next, we explore dynamic content caching in heterogeneous networks (HetNets) to minimize the AoI of dynamic contents being cached. We formulate the problem as a constrained Markov decision process (CMDP) and identify a threshold structure of the optimal solution. We design a deep deterministic policy gradient (DDPG) algorithm that exploits the threshold structure of the solution. Our proposed method achieves up to 30% reduction in AoI compared to a heuristic strategy. Finally, we propose the Disentangled REgion-of-interest Attention Map (DREAM) framework for privacy-aware region-of-interest (RoI) segmentation in edge inference systems. By incorporating attention mechanism and adversarial training, the DREAM architecture protects sensitive user information in segmentation tasks, reducing the eavesdropper accuracy by up to 38% without significantly compromising the segmentation performance.
|
| Genre | |
| Type | |
| Language |
eng
|
| Date Available |
2024-12-10
|
| Provider |
Vancouver : University of British Columbia Library
|
| Rights |
Attribution-NonCommercial-NoDerivatives 4.0 International
|
| DOI |
10.14288/1.0447451
|
| URI | |
| Degree | |
| Program | |
| Affiliation | |
| Degree Grantor |
University of British Columbia
|
| Graduation Date |
2025-05
|
| Campus | |
| Scholarly Level |
Graduate
|
| Rights URI | |
| Aggregated Source Repository |
DSpace
|
Item Media
Loading media...
Item Citations and Data
Permanent URL (DOI):
Copied to clipboard.Rights
Attribution-NonCommercial-NoDerivatives 4.0 International