Verifiable Distributed Computing via Multi-layer Data Integrity Auditing

Abstract

Distributed learning is regarded as a key solution to support edge computing in the era of digital prosperity. It significantly reduces the data transmission and privacy risks in centralized machine learning through local training with "data not leaving the edge". However, the quality of client data in distributed computing is often uneven, leading to inaccurate uploaded gradients and affecting the convergence and performance of the global model. To address the decline in model performance caused by untrusted client data, this paper proposes a federated learning admission and continuous verification mechanism based on data possession proofs for heterogeneous computing power. This mechanism will perform corresponding dataset verification based on the different computing power of clients. For strong devices, a complete S-PDP verification is executed, and the generated proof is signed with BLS to support multi-signature aggregation verification on the server side; for weak devices, a lightweight E-PDP is executed, only conducting a small number of samplings to reduce the overhead, and multiple rounds of verification are performed to enhance detection capabilities, and the proof is not signed to further reduce the burden on the edge side. Additionally, within the federated learning training cycle, this paper innovatively introduces a continuous random sampling verification mechanism to eliminate clients that fail to pass the proof for consecutive times, thereby continuously filtering out data anomalies or fake nodes while ensuring resource-friendliness, and enhancing the credibility of the federated training process and the robustness of the global model.