../

FedEmo: A Federated Learning Framework for Privacy-Preserving Emotion Detection From Handwriting on Consumer IoMT Devices

  • Some sort of model to recognize emotion from handwriting features
  • They train on the EMOTHAW dataset and achieve 92.65% accuracy when it is centralized
  • They achieve 87.3% accuracy under FL
  • They actually deployed this. They have commented on the bandwidth, packet loss etc…
  • They had to use the cloud to perform some of the task. Though it seems a bit vague how they came up with the latency metrics.
  • Under FL they split the dataset naively. They main differentiating factor in FL compared to distributed machine learning is the non-IIDness of the data. This cannot be achieved by splitting a large dataset.