Coded computing is currently being used for solving significant challenges in distributed computing systems, particularly in (decentralized) Federated Learning, the distributed form of ML. These challenges include particularly data privacy during distributed training, where coding is used to conceal information on the model and the training samples to active or passive attackers. However, coded computing is generally designed for exact recovery of the data, and works only for a limited class of functions and for computations on finite alphabets. This paper validates the Private Berrut Approximated Coded Computing with Secret Sharing (PBSS) as a solution to privacy concerns tailored for general FL settings. PBSS leverages Berrut Approximated Coded Computing (BACC) to implement a Secret Sharing protocol capable of computing non-linear functions while maintaining a balance between privacy, precision, and complexity. We integrate PBSS into a FL framework for image processing using two popular datasets (MNIST and CIFAR-10) and two different aggregation functions (FedAvg and FedMedian) with two different restrictions: dealing with honest but curious nodes and dealing with malicious nodes. We have performed different tests under these conditions to measure privacy, precision, and complexity to determine the viability of the trade-offs for realworld FL scenarios, concluding this approach offers promising results.