Enhancing Federated Learning in Consumer Electronics with Decoupled Knowledge Distillation against Data Poisoning

Firm data regulations worldwide limit Artificial Intelligence (AI) from collecting data for training models in consumer electronics, including AI-based medical equipment and smart home appliances. Federated Learning (FL) enables collaborative machine learning where clients train locally and a central server aggregates model parameters, eliminating the need for data centralization. However, FL is susceptible to data poisoning risks. Existing works typically concentrate on server aggregation, often overlooking the data poisoning impacts <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">on</sup> the honest clients' side. To this end, we present a DKDFL, which integrates Decoupled Knowledge Distillation(DKD) and data augmentation against poisoning during FL aggregation. Compared with traditional methods, the extensive experimental results demonstrate that the DKDFL achieves a maximum predictive accuracy advantage of 15.6% and an average accuracy leading of 6.57 %.