Sirvi Autor "Olapade, Mayowa Olaide, juhendaja" järgi

Nüüd näidatakse 1 - 1 1

Edge-intellekt käsu peale
(Tartu Ülikool, 2025) Kuchida, Reo; Flores Macario, Huber Raul, juhendaja; Olapade, Mayowa Olaide, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
The deployment of distributed machine learning (DML) at the edge has introduced a new generation of intelligent, real-time applications in domains such as autonomous vehicles, pervasive robotics, smart environments, and ambient sensing. However, managing such infrastructures remains inherently complex, particularly in decentralized, resource-constrained, and dynamically changing environments populated by heterogeneous devices. Existing solutions often rely on static configurations or require expert knowledge for setup, orchestration, and maintenance, thereby limiting their accessibility and scalability in real-world contexts. This thesis introduces \textit{Edge Intelligence on Command}, a novel architecture that employs large language models (LLMs) to initiate decentralized machine learning workflows via natural language intent. By combining prompt-engineered LLM agents with opportunistic device discovery and configuration, the system eliminates the need for specialized knowledge, empowering non-experts to deploy and manage ML tasks seamlessly. The architecture supports both federated and split learning, facilitating privacy-preserving, resource-efficient collaboration across diverse edge nodes. Beyond orchestration, the system addresses a critical challenge in DML deployments, which is ensuring the integrity of the learning process in the presence of abnormal device behavior. In particular, it focuses on poisoning attacks, where compromised or faulty edge nodes introduce corrupted data that can degrade model performance. In order to investigate this, a feasibility study conducted with a Raspberry Pi and thermal imaging reveals that poisoned training data induces detectable shifts in runtime behavior, including elevated temperature and increased CPU usage. These observations motivate the introduction of the Device Change Point Index (DCPI), a lightweight, decentralized anomaly detection mechanism based on native device metrics. Without relying on external hardware or centralized oversight, DCPI makes real-time trust assessment in edge learning environments. Taken together, this work demonstrates the feasibility and effectiveness of combining LLM-driven orchestration with runtime behavioral monitoring to enable secure, adaptive, and user-centric distributed intelligence at the edge. It contributes a practical step toward more autonomous and accessible edge AI systems.