Secure vertical federated learning based on feature disentanglement

•A new vertical federated learning framework (SVFL) preventing label owner from label inference attack.•A feature disentanglement based method to prevent participants who do not own labels from having too much information about the global model.•Participants who do not own labels cannot steal label information from the SVFL system, regardless of the percentage of data features they have.•SVFL makes the model more accurate due to the adversarial training involved in the feature disentanglement process.