To address the dual challenges of high communication overhead and privacy leakage that the mainstream federated learning algorithm FedAvg faces in collaborative sensitive financial risk control data, as well as the issue of model convergence degradation caused by clipping distortion in traditional gradient compression and encryption schemes, this paper proposes a Sparse Coordinate Encoding and Lossless Aggregation Federated Learning framework(SCE-LA-FL) . Firstly, a one-dimensional convolutional neural network is employed to automatically extract high-dimensional financial features. A preprocessing mechanism combining feature normalization, gradient quantization, and Top-K sparsification achieves significant compression of gradient magnitudes. Secondly, a lossless aggregation mechanism based on coordinate matching is designed to eliminate secondary distortion introduced during server-side aggregation by conventional bit-packing techniques. Finally, the framework integrates with additive homomorphic encryption(Paillier) to selectively encrypt only the retained sparse update values. Experiments on the LendingClub and MNIST datasets demonstrate that the proposed framework achieves a communication compression ratio of 1.25 to 9.98 times. Under integrated encryption, the area under the curve(AUC) for defending against member inference attacks drops to 48.4, the mean squared error(MSE) against model inversion attacks increases to 52.3, and the consistency rate of model extraction attacks decreases to 9.4%. This framework significantly reduces communication overhead while maintaining model convergence performance, and consistently defends against privacy attacks, offering valuable insights for financial credit risk prediction.