The paper introduces QFedAgent, a hybrid quantum-classical personalized federated learning framework designed for multi-agent activity recognition. It addresses the challenges of heterogeneous and non-IID multimodal sensor streams by integrating a variational quantum circuit fusion module that models accelerometer-gyroscope interactions through quantum state encoding.
- The approach utilizes only 72 quantum rotation parameters compared to 33K in classical multi-layer perceptron-based fusion, achieving approximately 10x total parameter reduction.
- Experiments on the OPPORTUNITY dataset under subject-based non-IID partitions demonstrate a mean test accuracy of 97.7%.
- The results confirm that parameter-efficient quantum fusion remains competitive with conventional federated baselines.
The framework offers a solution for privacy-sensitive robotic sensing applications by significantly lowering communication costs and parameter overhead while maintaining high accuracy.