Overview & Objective
Designed and evaluated a distributed Federated Learning (FL) platform at LINEACT Research Laboratory, benchmarking decentralized collaborative model training across a cluster of physical Raspberry Pi edge nodes and high-performance aggregation servers.
Key Engineering Highlights
Edge-to-Cloud Distributed Training
Deployed local training nodes on ARM-based single-board computers (Raspberry Pi), running decentralized gradient computations without transmitting raw private data.
Federated Aggregation & Non-IID Robustness
Implemented Federated Averaging (FedAvg) algorithms and evaluated convergence stability under severe non-IID (heterogeneous) data partitions across clients.
Optimized Communication Pipeline
Built lightweight network communication protocols using sockets and MQTT to minimize bandwidth overhead and transmission latency during parameter exchange.
Containerized Cluster Orchestration
Containerized workers and aggregation servers using Docker and Docker Swarm for reproducible multi-node experimentation and metrics collection.