Heterogeneous Federated Learning Frameworks for Balancing Job Completion Time and Model Accuracy

Federated Learning (FL) is a secure distributed learning paradigm, which enables potentially a large number of devices to collaboratively train a global model based on their local dataset. FL exhibits two distinctive features in job requirement and client participation, where FL jobs may have different training criteria, and clients possess diverse device capabilities and data characteristics. In order to capture such heterogeneities, this paper proposes a new FL framework, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Hca</i>, which aims to strike a balance between the job completion time and model accuracy. Specifically, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Hca</i> builds upon a number of innovations in the following three phases: 1) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">pre-estimation</i>: we first derive the optimal set of parameters used in training in terms of the number of training rounds, the number of iterations and the number of participating clients in each round; 2) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">client selection</i>: we design a novel device selection algorithm, which selects the most effective clients for participation based on both client historical contributions and data effectiveness; and 3) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">model aggregation</i>: we improve the classic FedAvg algorithm by integrating the utility in consecutive rounds as a weighted factor into aggregation computation. We further improve our model to consider a more realistic scenario, where multiple performance factors are difficult to quantify and coupled with each other, under the privacy protection requirement. We propose <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HcaRL</i>, a deep reinforcement learning (DRL) based smart client selection and model aggregation algorithm, to solve the problem in the above scenario. To evaluate the performance and effectiveness of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Hca</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HcaRL</i>, we conduct theoretical analysis and testbed experiments over an FL platform FAVOR. Extensive results show that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Hca</i> can improve the job completion time by up to 34% and the model accuracy by up to 9.1%, and can reduce the number of communication rounds required in FL by up to 75% compared with two state-of-the-art FL frameworks. In addition, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HcaRL</i> has a further performance improvement over <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Hca</i>, including a 5% increase in model accuracy and a 23.6% reduction in job completion time.