Federated transfer learning in smart homes pattern recognition for privacy preservation

Abstract

Smart homes collect continuous sensor data about how people live, including when they sleep, cook, work, and move around the house. Training AI models on this data has traditionally required sending it to a central server, which creates serious privacy risks and conflicts with laws like the General Data Protection Regulation. This thesis builds and tests a Federated Transfer Learning framework that trains a shared AI model across ten smart homes without any raw data ever leaving the local device.

The system uses a two-layer Long Short-Term Memory network trained through the Federated Averaging algorithm across ten distinct household types. The model is split into two parts. The shared body learns general activity patterns from all ten homes and is the only part transmitted between clients and the server. The private head stays on each device throughout all 60 rounds of training, is never transmitted to the server, and adapts privately to each household's specific activity distribution.

Testing showed that eight of ten clients' private heads achieved 100.00% macro F1 on their own local validation data, Client 4 achieved 99.63%, and Client 10 achieved 99.67%, confirming that private head training across 60 rounds produces strong household-specific classification without any head information leaving the local device. On a shared test set evaluated as the average across all ten client models, the FTL-LSTM framework achieved 90.86% macro F1. This lower figure reflects the combined effect of head privacy and household specialization: nine of the ten private heads were trained on households other than the test home. Standard FedAvg, which transmits full model weights including head parameters every round across the same 60 rounds, achieved 99.61% on the same shared test set using a single globally averaged head. The 3.7% communication saving from body-only transmission was confirmed with FTL-LSTM transmitting 575.39 MB compared to 597.41 MB for standard FedAvg over identical 60-round durations. In the cold start scenario evaluated on 865 windows from Home 11, the pretrained body provided a transfer advantage in all four labeled sample conditions: +9.80 percentage points at 10 samples, +8.33 at 20 samples, +12.18 at 50 samples, and +8.31 at 100 samples, with peak FTL performance of 52.29% at 100 samples. These results show that body only head private federation produces near perfect local performance for each household while carrying a measurable difference on shared external test evaluations, a difference that reflects both head privacy and household specialization rather than privacy cost alone.