Differential Privacy in Federated Learning for Collaborative Radio Map Construction and Environment Sensing

Collaborative radio map facilitates precise and secure monitoring and management of wireless communication environment by encouraging collective participation and information sharing. However, the collaborative construction of radio map requires a substantial amount of data and raises significant concerns about user privacy. This paper proposes a federated learning strategy with differential privacy, aiming to safeguard user location privacy for collaborative radio map and virtual environment construction. It addresses inherent heterogeneity challenge in local sensing data by designing a scaling matrix for the local gradients, effectively mitigating imbalances among different users. In addition, this paper proposes an adaptive heterogeneous noise design for differential privacy to optimize the allocation of noise added to local gradients, seeking a near-optimal balance between preserving location privacy and maintaining radio map construction performance. Experiments demonstrate that the proposed method fulfills a remarkable around 200% improvement in uncertainty of external adversary localization, accompanied by only 0.2 dB deviation in the construction accuracy of the radio map.