Additive manufacturing (AM) has significantly transformed production processes by enabling innovative design flexibility and rapid prototyping. Despite these advantages, maintaining consistent product quality in AM remains challenging, with a notable defect rate that impacts production costs and efficiency. Traditional methods of AM process anomaly detection may require aggregating data collected from different sources in a centralized manner to train a reliable machine learning model, raising privacy concerns. This study proposes a federated learning (FL) framework combined with convolutional neural networks (CNN) to address these issues. FL enables decentralized model training across multiple clients, ensuring data privacy by not sharing raw data. Experimental results demonstrate that the FL approach outperforms Individual learning (IL) and centralized learning (CL) in terms of F1 score, Precision and Recall. The FL model achieved an F1 score of 0.897, compared to 0.876 for CL, 0.852 and 0.826 for two IL, indicating superior AM process anomaly defect detection performance while maintaining data privacy. This research highlights the potential of FL in enhancing defect detection in AM processes, providing a robust solution for privacy-preserving process-defect modelling.