The rapid growth of Internet of Things (IoT) devices has expanded the attack surface of modern networks while introducing severe constraints in computation, energy, and communication. Traditional intrusion detection systems (IDS) are either centralized, incurring high privacy and bandwidth costs, or too heavy for low-power IoT devices. Federated learning (FL) provides a privacy preserving alternative but suffers from high communication overhead, heterogeneous client performance, and vulnerability to adversarial updates. This paper presents L-FIDS, a lightweight intrusion detection system for IoT devices. It uses a small 1D-CNN model and only allows trusted devices to take part in training. L-FIDS keeps user data private, reduces communication costs, and can still detect attacks well. We use the BoT-IoT and TON_IoT datasets for the experiments, and the results show that L-FIDS achieves over 92% accuracy, reduces communication by more than 70%, and keeps accuracy above 91% even when some devices try to send harmful updates.