Deep learning enabled semantic communications and data trading

• Main Author: Liew, Zi Qin
• Other Authors: Dusit Niyato
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/176768
• Institution: Nanyang Technological University

The semantic communication system enables wireless devices to communicate effectively with the semantic meaning of the data. The ability of semantic communication systems to filter the relevant information for effective data communications is envisioned to be one of the key strategies to improve the efficiency and sustainability of the next-generation networks. However, most of the resource allocation frameworks and incentive mechanism designs for conventional communication systems aim to maximize the throughput of the networks without considering the semantic importance of the bit flow. In this thesis, we study and design system models and incentive mechanisms for semantic communication systems. In the first study, a hierarchical trading system is designed to support the semantic model and semantic information trading jointly. The incentive mechanism we propose serves to optimize the revenue of semantic model providers in the semantic model trading sector, effectively encouraging model providers to actively contribute to the development of semantic communication systems. For semantic information trading, our auction approach can facilitate transactions among multiple sellers and buyers of semantic information. It ensures individual rationality, incentive compatibility, and budget balance, enabling participants to attain higher utilities compared to the baseline method. In the second study, we propose a semantic communication system where data for building Digital Twins (DTs) is encoded as semantic information, such as scene graphs. Virtual Service Providers (VSPs) can use these graphs to build or update their DTs. To facilitate data trading between VSPs and edge nodes, we introduce a combinatorial auction-based incentive mechanism where we model the valuation of semantic data based on relatedness, freshness, and semantic values. We measure the relatedness based on the similarity between targeted objects of VSPs and the objects present in the real-world data. We use the age of information (AoI) to quantify data freshness and evaluate the semantic value using the information richness of the scene graphs. Simulation results show that our system can maintain the model's mean recall performance using only 12.5% of features during transmission. Our mechanism increases the average update frequency effectively. The theoretical analysis proves that the proposed mechanism guarantees incentive compatibility and individual rationality. In the third study, we propose artificial intelligence-generated bidding (AIGB) to optimize the bidding strategies for artificial intelligence-generated content (AIGC). AIGB model uses the reinforcement learning model to generate bids for the services by learning from the historical data and environment states such as remaining budget, budget consumption rate, and quality of the won services. To obtain quality AIGC service, we propose a semantic aware reward function for the AIGB model. The proposed model is tested with a real-world dataset and experiments show that our model outperforms the classical bidding mechanism in terms of the number of services won and the similarity score. Finally, we conclude the thesis and highlight the future research directions.