Table 1.
Qualitative Comparison between IVA-FL and Existing Adaptive DP Paradigms. Unlike existing methods that prioritize convergence or user preference, IVA-FL is the first to prioritize the preservation of minority class signals based on data semantics.
Fig 1.
The overall framework of IVA-FL.
Table 2.
Core performance comparison of IVA-FL and baseline methods on two imbalanced datasets.
Fig 2.
Trend of AUC performance with privacy budget (ε).
Fig 3.
Trend of Recall with Class Imbalance (Fixed ε = 4).
Table 3.
Financial Risk Assessment Metrics Comparison. KS Statistic measures discrimination capability (Higher is better), and Brier Score measures probability calibration (Lower is better). Data is based on the Credit Card Fraud Detection dataset.
Fig 4.
The KS curve measures the maximum difference between the cumulative distributions of positive (fraud) and negative (normal) samples. A higher curve peak indicates better risk discrimination capability. IVA-FL demonstrates a significantly higher KS Statistic (0.742) compared to baselines.
Fig 5.
Robustness Analysis against Data Heterogeneity (Non-IID).
The experiment evaluates model Recall under varying Dirichlet parameters α. A smaller α (e.g., 0.1) indicates extreme data heterogeneity. While baseline methods degrade significantly in extreme Non-IID settings, IVA-FL demonstrates strong resilience, maintaining high utility.
Table 4.
IVA-FL core mechanism ablation study results (ε = 4).
Fig 6.
Comprehensive Ablation Study: Model Performance Comparison (ε = 4).
Fig 7.
Scalability Analysis: Final Model Performance vs. Number of Clients.
Fig 8.
Scalability Analysis: Convergence Speed vs. Number of Clients.
Table 5.
Comparison of Computational Overhead and Resource Consumption per Communication Round. The experiment was conducted on the Credit Card Fraud Detection dataset with batch size 64.
Fig 9.
Computation Overhead and Efficiency Analysis.
Fig 10.
Parameter Sensitivity Analysis.
(a) Impact of Clipping Modulation Factor λ on Recall and AUC. (b) Impact of Smoothing Decay Factor γ on Recall and AUC. The shaded areas represent the standard deviation across 5 independent runs. The results indicate that the model achieves optimal performance around and
, exhibiting a stable inverted U-shaped trend.
Table 6.
Model performance comparison in a simulated SCF feature-heterogeneous scenario.
Fig 11.
Performance in Simulated Heterogeneous SCF Scenario.
Table 7.
Evolution of Model AUC Performance in a Dynamic Market Environment.
Fig 12.
Adapted Performance Comparison After Concept Drift.