A Multi-Layer Auditable Vertical Federated Learning Prototype for Power Equipment Supply Chains: Reproducibility, Robustness, and Privacy-Boundary Evaluation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Transformer lifecycle data across organizations are typically vertically partitioned among material suppliers, manufacturers, logistics service providers, testing agencies, and operation and maintenance units. This study presents a reproducible multi-layer vertical federated learning (VFL) prototype that integrates salted-hash identifier matching, additive sharing of local score vectors over finite fields, and a local public key infrastructure with a signature-based audit verification mechanism. A deterministic synthetic dataset is first constructed, comprising 5200 aligned records and 31 predictor variables, which are partitioned among five participants with varying numbers of features per participant. Second, across five validation runs, the VFL models under both the standard block-wise and score-sharing paths achieved an AUC of 0.8825 ± 0.0119, an F1 score of 0.7367 ± 0.0249, and an accuracy of 0.8102 ± 0.0183 on the test set. The classification results of both paths were fully consistent with the centralized gradient-descent logistic regression baseline. Notably, the score-sharing path exhibited a maximum log-odds deviation of only 2.22 × 10−8 on the test set, with no prediction discrepancies observed. Third, across 10 independently generated synthetic populations, the nonlinear output mechanism highlights the limitations of linear models: the AUC of vertical federated learning (VFL) drops to 0.6457 ± 0.0171, while Extra Trees and HistGradientBoosting achieve 0.7731 ± 0.0149 and 0.7743 ± 0.0139, respectively. Finally, in a separate residual-sharing diagnostic test, when 1 to 4 participants jointly shared the residuals, the label inference AUC remained around 0.499–0.500; however, when all five participants shared or plaintext residuals were used, the labels could be fully recovered. Both simple membership inference diagnostic tests yielded results close to random. The local signature log verifier rejected all 700 injected faults and accepted the 400 clean control log events. These results validate the feasibility of numerical reproducibility and local audit functionality under synthetic data and single-process conditions, yet they are insufficient to demonstrate end-to-end label privacy protection, malicious security, effectiveness on real data, or real-time ledger performance.