Federated Learning (FL) allows devices, like smartphones and computers, to collaboratively train a predictive model while keeping their data locally stored. This approach enables machine learning without relying on cloud storage. In FL, clients share computed gradients or model parameters instead of raw data, but privacy concerns arise when local information can still be inferred from these outputs. To address this, clients often add artificial noise or encryption to their updates. However, even with noise, gradients can be reverse-engineered, potentially exposing sensitive data. Recent advancements focus on sharing model parameters rather than local data or gradients, enhancing privacy through encryption and additional noise. This paper introduces a federated learning algorithm that utilizes coded local gradients during training and exchanges coded combinations of learned model parameters, thereby ensuring privacy without encryption. By injecting surplus noise into these proxy variables, the algorithm establishes a stronger privacy framework. The author presents two variants of the algorithm: a general version and a specific one that operates without surplus noise. He demonstrates the algorithms’ convergence and provides insights into improving learning rates based on coding schemes aligned with the data’s characteristics. Two encryption-free implementations, using fixed and random coding matrices, yielding promising simulation results are introduced. Both are viewed from both a federated optimization and a federated machine learning perspective.