In this paper, we propose a novel federated learning of random oblique stumps (FL-ROS) for handling the ImageNet challenge having 1,281,167 images and 1,000 classes. Our FL-ROS algorithm trains an ensemble random oblique stumps on Raspberry Pi Zeros (RPi Zeros) without exchanging data among RPi Zeros, to classify the ImageNet dataset. The multi-class Proximal Support Vector Machines (MC-PSVM) uses the One-Versus-All (OVA) multi-class strategy and the under-sampling technique for independently learning random oblique stumps from the local training subset stored on RPi Zeros. The empirical test results on the ImageNet dataset show that our FL-ROS algorithm with 4 RPi Zeros (Quad-core 64-bit ARM Cortex-A53 processor clocked at 1GHz and 512MB RAM) is faster and more accurate than the state-of-the-art SVM algorithms run on a PC (Intel(R) Core i7-4790 CPU, 3.6 GHz, 4 cores, 32GB RAM).