Federated learning, as a universal machine learning paradigm for addressing data silos and data privacy, enables collaborative model training without sharing raw data. However, existing research indicates that the transmission of plaintext gradients still faces privacy leakage risks. Current gradient obfuscation methods are effective against reconstruction attacks but fail to defend against label inference, while encryption methods rely on third parties and incur high overhead. In this study, we propose the FedSND framework: by decomposing gradients using SVD and encrypting them with decentralized random orthogonal matrices, combined with threshold secret sharing incorporating symmetric noise, we achieve full-process privacy protection without third-party trust. Experiments on the MNIST/CIFAR-10 datasets demonstrate that FedSND reduces the accuracy of label inference attacks by 99.9% and decreases the PSNR of reconstructed data to 6.51 dB and 6.37 dB, respectively, significantly outperforming traditional defense methods.