Unbiased federated learning model for skin lesion classification

Federated Learning

Multimodal images

Domain Adaptation

Debiased Method

TABE: use custom loss which minimize the main classification loss, and train additional classifier using confusion loss between additional class label and uniform distribution with the same feature representation

Additional fusion branch

Multi-scale Fully-shared Fusion Network Multi-scale fusion structure combines deep and shallow features within individual modalities, and Dermo-Clinical Block (DCB) integrates the feature maps from dermoscopic images and clinical images

Domain-Adversarial Training of Neural Networks (Gradient Reversal) Train a domain classifier and multiplies the gradient by a certain negative constant during the backpropagation-based training on feature extractor layer's weights.

Late fusion

Performance Measurement

Accuracy benefit of FL: accuracy of (FL model - local model) on local test set

Fairness benefit of FL: fairness of (FL model - local model) on local test set. Fairness is measured using equalized odds (Hardt et al., 2016) and demographic parity (Dwork et al., 2012)

Communication efficiency

Local updating

Compression schemes

Decentralized training

Systems heterogeneity

Topology

Statistical heterogeneity

Asynchronous communication

Active sampling

Fault tolerance

Modeling heterogeneous data

Bias reduction

Reveal bias

Fairness

Refine client update

Regularize the local objective function

Personalized FL

click to edit