Training criteria: Accuracy and loss values (15 epochs). The benchmark in these experiments is a model that was trained using a centralized dataset of 10,000 images of the same architecture for 15 epochs. The benchmark model reached a test accuracy of 45.06% and loss of 1.41.
Results: Overall, BL and FL algorithms can provide acceptable utility compared to centralized training. However, FL requires almost twice the time of BL to converge to a similar accuracy. The execution time of FL is always higher than BL when both are run in a real cloud setup using our framework. The following figures illustrate the accuracy/epoch in BL, FL, and the benchmark. Note that BL converges to a level equivalent to the benchmark.
It is also evident from the figure below that BL reaches a loss value that is almost identical to the benchmark. It is, therefore, safe to assume that BL approaches can converge to identical results as a centralized learning approach, where the data and the model are on the same device.