Knee osteoarthritis (KOA) is a degenerative joint disease commonly assessed from radiographic images using the Kellgren–Lawrence (KL) grading system. This study proposes a lightweight federated learning framework for automatic KOA severity classification by combining a LightCNN backbone with the Convolutional Block Attention Module (CBAM). The model was evaluated on the Knee Osteoarthritis Severity Grading Dataset, comprising 8,260 X-ray images across five KL grades, under both centralized and federated learning settings. In the federated setting, three clients collaboratively trained the model using Federated Averaging (FedAvg) for up to 50 communication rounds without sharing raw images. The centralized and federated models achieved accuracies of 64.55% and 61.35%, with macro F1-scores of 64.86% and 60.88%, respectively, corre- sponding to a 3.20 percentage-point difference in accuracy. Grade 1 remained the most challenging category, while Grad-CAM showed that the model primarily attended to relevant regions around the knee joint. These results indicate that LightCNN- CBAM can retain competitive classification performance in afederated setting while providing a compact and interpretable framework for collaborative KOA assessment.