Table 10 Unified Comparative Evaluation of MedShieldFL vs. Baselines.
Method | Acc. (%) | G1 | G2 | G3 | Train/Infer Time | Privacy & IIoT Suitability |
|---|---|---|---|---|---|---|
Traditional ML (SVM, RF)21 | 84.7 | – | – | – | 8 s/1.2ms | Low privacy; weak IIoT fit |
Centralized DL (CNN, U-Net)36 | 90.3 | – | – | – | 28 s/2.4ms | Moderate privacy risk; requires central storage |
Centralized (Real)32 | 97.08 | 94.4 | 97.7 | 97.7 | –/– | Not privacy-preserving |
Centralized (Synthetic)33 | 99.35 | 100.0 | 98.5 | 98.6 | –/– | Unrealistic; lacks real data variability |
Centralized (Mixed)12 | 98.37 | 97.3 | 99.2 | 98.4 | 28 s/2.4ms | High privacy risk; high performance |
FL (FedAvg)34 | 95.72 | 91.2 | 95.4 | 96.2 | 32 s/2.1ms | Moderate privacy; deployable on edge |
FL + DCGAN7 | 96.85 | 93.6 | 96.7 | 97.3 | 50 s/2.8ms | Improved generalization; GAN instability risk |
FL + HE26 | 96.12 | 92.5 | 95.8 | 96.1 | 45 s/2.6ms | Strong privacy; reasonable cost |
FL + GAN6 | 93.6 | – | – | – | 50 s/2.8ms | GAN-augmented; good for IIoT |
MedShieldFL (Ours) | 98.37 | 97.3 | 99.2 | 98.4 | 54 s/2.5ms | HE + GAN; Excellent privacy, robustness, and IIoT deployment suitability |
