Outsourced Verifiable Privacy-preserving Federated Learning via Sensitive Samples

Federated Learning (FL) based on Differential Privacy (DP) has become a popular method to enhance data privacy by protecting client data contributions during training. However, existing solutions typically assume a uniform privacy budget for all records, providing one-size-fits-all solutions that may fail to meet the diverse privacy requirements of individual records. Furthermore, FL vulnerable to model integrity attacks when resource-constrained users outsource training processes to cloud services. In this paper, We propose OVeriFL, an outsourced verifiable privacy-preserving FL scheme that integrates Personalized Differential Privacy (PDP) for adaptive privacy protection and uses sensitive samples for model integrity verification. Specifically, we define sensitive samples as a set of imperceptibly transformed inputs that are highly sensitive to model parameter changes. Even minor modifications to the model can be clearly reflected in the model outputs. Experimental results show that OVeriFL not only reliably detects model integrity attacks in cloud-hosted environments but also improves model utility by mitigating catastrophic forgetting and reducing accumulated noise through its PDP mechanism with the Poisson sampling procedure.