Federated Machine Learning (FML) enables multiple entities to collaboratively train a shared ML model while keeping their local datasets private. By combining data and knowledge across participants, FML can produce better models than those trained on isolated datasets. The preprocessing of training data, such as scaling numerical features and encoding categorical variables, is essential to ensure data uniformity and thus improve model quality. In horizontal FML systems, participants have data with the same features, but own different samples. Achieving a uniform preprocessing among participants usually requires revealing statistics about training data or even the exchange of sensitive categorical values, which undermines privacy. To avoid it, privacy-preserving preprocessing methods are necessary. This work addresses the secure scaling of numerical features and the consistent encoding of categorical data in FML systems. We propose a privacy-preserving approach for Z-Score normalization that allows participants to scale numerical features collaboratively without revealing local statistics. We evaluate eight methods for a uniform encoding of categorical values across the system and assess their impact on model quality, dimensionality, scalability, and privacy. Our experimental evaluation on five datasets demonstrates that system-wide consistent preprocessing improves model performance compared to isolated approaches. However, we observe a trade-off between privacy preservation and efficiency. Based on our findings, we provide practical recommendations for selecting suitable encoding strategies depending on the use case, aiming to balance privacy, data consistency, and model performance in FML settings.