Electric vehicles are increasingly integrated into the charging infrastructure, making them more vulnerable to cyberattacks. In this paper, a privacy-sensitive, scalable intrusion-detection system proposes applying federated learning to a distributed multi-layer perceptron (MLP) classifier implemented on five charging-site clients. Each client preprocesses local features, selects them, and updates the model, which is then averaged on FedAvg at the server. We test the system on a multiclass EV-charging system (CICEVSE2024) and demonstrate that the federated system can effectively detect various attack types, such as large-scale network floods, achieving a best test accuracy of 92 percent, good per-class classification, and high tolerance to extreme class imbalance. The results show that federated learning enables effective collaboration among sites with conflicting charging needs, allowing them to work together without sharing raw telemetry, thereby reducing privacy risks and communication overhead. The proposed framework can also be implemented in a decentralized manner across main mobility networks and provides methods for addressing heterogeneity and imbalance in real-world IDS tasks.