Skip to main content
. 2025 Aug 22;15:30915. doi: 10.1038/s41598-025-16604-2

Table 1.

Conceptual comparison of the proposed method and three representative SOTA approaches.

Criterion Proposed (Ours) FedEntropy EDS-FL SER
Design paradigm Regularised representation optimisation Entropy-constrained quantisation Task-oriented distillation Stochastic latent encoding
Entropy reduction potential High (explicit regularisation of representation entropy) Moderate (adaptive to local entropy) Moderate (guided reduction via distillation) Low (uncontrolled stochastic variability)
Semantic fidelity High (controlled latent structure) Moderate (depends on encoder configuration) Moderate (teacher-guided relevance) Low (sample-level uncertainty dominates)
Task adaptivity High (task-aware training objectives) Moderate (manual hyperparameter tuning) Moderate (aligned via teacher) Low (static latent sampling)
Generalisation capability High (validated for unseen classes/tasks) Low (encoder-specific generalisation) Moderate (task-bound distillation) Low (limited transferability)
Compression adaptability High (context-aware encoding) Moderate (entropy-profile driven) Moderate (fixed distillation targets) High (sampling enables variability, less control)
Latent space consistency High (stabilised via regularisation) Moderate Low (inter-client feature divergence) Low (sampling noise)
Interpretability Moderate–High (structured latent space) Low Moderate Low
Scalability High (50 + clients, multiple tasks supported) Moderate (tested on small-scale setups) Low (distillation overhead) Moderate (client-specific tuning needed)
Communication efficiency High (learned, compact, and robust encoding) Moderate (entropy heuristics) Low (additional teacher–client exchange) Moderate (bit-level control, low robustness)
Training complexity Moderate (no external modules) Moderate High (teacher synchronisation required) Moderate–High (training instability)
Robustness to data heterogeneity High (validated under non-IID distributions) Moderate (entropy adapts partially) Low (sensitive to class imbalance) Low (amplified by stochasticity)
Inference suitability for edge devices High (low latency and model size) Moderate Low (teacher model size dominates) Moderate (requires sampling overhead)