16 October 2024 in IEEE Access
Large Language Models (LLMs) empowered agents are effective across various tasks and demonstrate outstanding performance, which can be further enhanced through collaboration with multiple LLM agents. The current approaches for collaboration with multiple LLM agents are static approaches, which adopt a fixed set of agents to interact with each other. However, these approaches suffer from a significant limitation: multi-agent collaboration depends on the assumption that all participants know each other in a local closed environment, can find each other and direct communication, and will act with integrity. To address these challenges, this paper proposes DeCoAgent, a novel framework for decentralized autonomous collaboration between LLMs empowered agents based on smart contracts. This framework enables decentralized autonomous collaboration between LLM agents, allowing them to register themselves, discover the capabilities of other agents, and assign tasks on the platform. LLMs can convert natural language descriptions from human and LLM agent users into smart contract calls, enabling agents to interact with humans, the blockchain, and other agents to achieve automation. This paper implements the platform based on OpenAI and Ethereum, demonstrating the practical feasibility of this approach. The proposed framework has broader applications, including supply chain management, manufacturing, crowdsourcing, and complementing other existing multi-agent collaborations. This framework is open source on GitHub. Please visit the repository at https://github.com/AnanKing/DeCoAgent.
9 January 2025 in IEEE Access
Payment channel networks (PCNs) are a viable solution to the issue of blockchain scalability by offering off-chain transactions between two untrusted peers, without committing each transaction to the blockchain. The payment channel cannot be used when its deposit is depleted, which hinders the payment completion, such as it is overly utilized in one direction. However, they cannot migrate the deposit from one channel to another to facilitate payment completion when a depletion in the channel deposit arises. This limitation arises from the basic design assumption of payment channels, i.e., each channel is independent from others. To overcome this limitation, this paper proposes PaySwitch, inspired by the traditional network switch, to enable connecting multiple payment channels through one PaySwitch smart contract. PaySwitch is designed to enable any peer in the PaySwitch to dynamically reallocate the deposits in their payment channels while maintaining the independence of each other to ensure the trustworthiness of the channel. This paper presents the security definition of PaySwitch and formally proves its security using the UC-framework. To demonstrate the feasibility of PaySwitch, this paper applies it to several routing algorithms. The experimental results show that PaySwitch is capable of improving the transaction success rate and throughput compared with the traditional payment channel based approaches.
6 November 2023 in International Journal of Network Management
16 October 2026 in Expert Systems with Applications
7 August 2026 in SSRN Electronic Journal
Small and fast unmanned aerial vehicle (UAV) detection in real-world low-altitude visual signal processing scenarios remains challenging due to weak target texture, large scale variation, complex background interference, and motion-induced appearance degradation, which often lead to missed detections and unstable localization. To address these challenges, this paper proposes LME-YOLO, a lightweight multi-scale enhanced detection framework for robust small and fast UAV detection. The proposed framework integrates C3k2_WTConv to enhance high-frequency detail extraction for weak and motion-degraded targets, ADown to reduce information loss during downsampling, a binary-gated multi-scale attention fusion module to strengthen small-target representation in high-resolution features, and a center--scale regularized IoU--NWD fusion loss to improve localization robustness for tiny UAVs. Compared with state-of-the-art lightweight detectors on the DUT Anti-UAV dataset, our method achieves 2.95% higher mAP@0.5 and 0.92% higher mAP@0.5:0.95, while reducing computational complexity by 24% in FLOPs.