I am also looking forward to the federated learning with mobile devices and orchestration server. From my resent conversation with a ML compiler programmer, the tflite model runs on the mobile device are compiled, it would be difficult to decompile the tflite model to get the tf ml code back. Thus, it would be really difficult to retrain or using transfer-learning to modify the tflite model. Currently, i am using java/kotlin to build additional layer to the tflite model codes, so that the additional layer can be retrained on mobiel device. But still i have an issue to get the weights from mobile device aggregated, since it doesn’t really converge to an equilibrium.