FedNet: Federated implementation of neural networks for facial expression recognition

Abstract

Overfitting is a significant obstacle in classification tasks like Facial Expression Recognition. Despite repeated attempts to negate the effects of overfitting by various researchers, the issue persists. This paper introduces FedNet, a Neural Network architecture inspired by the federated aggregation and averaging technique used in Federated optimization, which proposes a method to learn from data situated separately in edge devices. Our study proposes dividing training data into multiple shards and training each data shard individually using the same neural network, then aggregating and averaging each model's parameters after every few epochs. Training multiple models on different data shards and averaging the learned parameters allows the model to learn from the entire dataset while avoiding overfitting on a particular data shard. Convolutional Neural Networks were implemented using the Extended Cohn Kanade and the FER-2013 dataset to conduct this study. Our federated averaging-based implementation of CNNs achieved 99.1% accuracy and 100% accuracy for 8 and 7 emotion classifications, respectively, on CK+, beating the benchmarks for this dataset. It also achieved 65.6% accuracy on FER-2013 without using any transfer learning or data augmentation. Our model shows significantly better resistance against overfitting, resulting in better generalization compared to the other existing methods.