Publicação
A federated learning framework for the next-generation machine learning systems
| Resumo: | The end of Moore's Law aligned with rising concerns about data privacy is forcing machine learning (ML) to shift from the cloud to the deep edge, near to the data source. In the next-generation ML systems, the inference and part of the training process will be performed right on the edge, while the cloud will be responsible for major ML model updates. This new computing paradigm, referred to by academia and industry researchers as federated learning, alleviates the cloud and network infrastructure while increasing data privacy. Recent advances have made it possible to efficiently execute the inference pass of quantized artificial neural networks on Arm Cortex-M and RISC-V (RV32IMCXpulp) microcontroller units (MCUs). Nevertheless, the training is still confined to the cloud, imposing the transaction of high volumes of private data over a network. To tackle this issue, this MSc thesis makes the first attempt to run a decentralized training in Arm Cortex-M MCUs. To port part of the training process to the deep edge is proposed L-SGD, a lightweight version of the stochastic gradient descent optimized for maximum speed and minimal memory footprint on Arm Cortex-M MCUs. The L-SGD is 16.35x faster than the TensorFlow solution while registering a memory footprint reduction of 13.72%. This comes at the cost of a negligible accuracy drop of only 0.12%. To merge local model updates returned by edge devices this MSc thesis proposes R-FedAvg, an implementation of the FedAvg algorithm that reduces the impact of faulty model updates returned by malicious devices. |
|---|---|
| Autores principais: | Costa, Diogo André Veiga |
| Assunto: | Federated learning Machine learning Artificial neural networks Artificial intelligence Machine learning algorithms Intelligent systems Internet of things Arm Cortex-M Treino federativo Redes neuronais arificiais Inteligência artificial Algoritmos de machine learning Sistemas inteligentes Internet das coisas |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | The end of Moore's Law aligned with rising concerns about data privacy is forcing machine learning (ML) to shift from the cloud to the deep edge, near to the data source. In the next-generation ML systems, the inference and part of the training process will be performed right on the edge, while the cloud will be responsible for major ML model updates. This new computing paradigm, referred to by academia and industry researchers as federated learning, alleviates the cloud and network infrastructure while increasing data privacy. Recent advances have made it possible to efficiently execute the inference pass of quantized artificial neural networks on Arm Cortex-M and RISC-V (RV32IMCXpulp) microcontroller units (MCUs). Nevertheless, the training is still confined to the cloud, imposing the transaction of high volumes of private data over a network. To tackle this issue, this MSc thesis makes the first attempt to run a decentralized training in Arm Cortex-M MCUs. To port part of the training process to the deep edge is proposed L-SGD, a lightweight version of the stochastic gradient descent optimized for maximum speed and minimal memory footprint on Arm Cortex-M MCUs. The L-SGD is 16.35x faster than the TensorFlow solution while registering a memory footprint reduction of 13.72%. This comes at the cost of a negligible accuracy drop of only 0.12%. To merge local model updates returned by edge devices this MSc thesis proposes R-FedAvg, an implementation of the FedAvg algorithm that reduces the impact of faulty model updates returned by malicious devices. |
|---|