Table 5. Comparison of affect recognition studies using federated learning (FL) and privacy-preserving approaches.^a.

Study	Dataset	FL algorithm	Data split	Accuracy
Almadhor et al [39]	WESADb	FedAvgc+logistic regression	N/Ad	86.82
Fauzi et al [40]	WESAD	FedAvg+DNNe network	N/A	85.75
Can and Ersoy [15]	Private dataset	FedAvg+MLPf	N/A	88.55
Lee et al [29]	WESAD	FedAvg+MLP	LOOCVg	75.00
Our previous study [38]	WESAD	FedAvg+1 DCNNh+DPi	5-fold CVj	90.00
This study	VERBIOk	FedAvg+1 DCNN+DP	5-fold CV	88.67

^aThe table reports the main approaches have been applied for stress detection by using a physiological dataset.

^bWESAD: wearable stress and affect detection.

^cFedAvg: federated averaging.

^dN/A: not applicable.

^eDNN: deep neural network.

^fMLP: multilayer perceptron.

^gLOOCV: leave-one-out cross-validation.

^hDCNN: deep convolutional neural network.

ⁱDP: differential privacy.

^jCV: cross-validation.

^kVERBIO: virtual environment for real-time biometric interaction and observation.