Toward Privacy-Preserving and Error-Tolerant Wireless Federated Learning: Fixed-Point Model Aggregation With Differential Privacy Guarantees

This paper presents a novel approach for wireless federated learning (WFL) that, for the first time, enables the aggregation of local models with mild to moderate errors under practical communication settings, which has to date been prevented by floating-point standards, e.g., IEEE binary32, and encryption. Specifically, we propose a new conversion from floating-point local models to fixed-point models on a layer basis, eliminating the need to transmit error-intolerant sign and exponent bits of floating-point numbers while accommodating variations in model layer widths and magnitudes. We also quantify how bit errors in the ciphertext affect the plaintext when symmetric encryption is employed for local model uploading, e.g., under Rayleigh, Rician, and Nakagami-m fading channels. Notably, these bit errors are leveraged to enhance the privacy of local models. We interpret the local model transmission process as a (λ, ϵ)-Rényi Differential Privacy (DP) mechanism, where bit errors induced by noisy channels, controlled via transmit powers, and exacerbated by decryption act as DP perturbations. Experiments show the superiority of the new WFL to the status quo with higher training accuracy and lower communication overhead.