EcoFair: Trustworthy and Energy-Aware Routing for Privacy-Preserving Vertically Partitioned Medical Inference

Abstract

Privacy-preserving medical inference must balance data locality, diagnostic reliability, and deployment efficiency. This paper presents EcoFair, a simulated vertically partitioned inference framework for dermatological diagnosis in which raw image and tabular data remain local and only modality-specific embeddings are transmitted for server-side multimodal fusion. EcoFair introduces a lightweight-first routing mechanism that selectively activates a heavier image encoder when local uncertainty or metadata-derived clinical risk indicates that additional computation is warranted. The routing decision combines predictive uncertainty, a safe--danger probability gap, and a tabular neurosymbolic risk score derived from patient age and lesion localisation. Experiments on three dermatology benchmarks show that EcoFair can substantially reduce edge-side inference energy in representative model pairings while remaining competitive in classification performance. The results further indicate that selective routing can improve subgroup-sensitive malignant-case behaviour in representative settings without modifying the global training objective. These findings position EcoFair as a practical framework for privacy-preserving and energy-aware medical inference under edge deployment constraints.

Figures (4)

• Figure 1: Architecture of the EcoFair framework in a simulated Edge-to-Edge-to-Cloud vertically partitioned inference setting. The image client (Party A1) processes raw image input locally using a lightweight encoder and computes routing signals from the local predictive distribution. Depending on the routing decision, the heavier image encoder may also be activated on the same client. The tabular client (Party A2) processes patient metadata locally to produce a tabular embedding and risk score. Only modality-specific embeddings are transmitted to the server (Party B), where multimodal fusion produces the final diagnostic prediction.
• Figure 2: Empirical malignancy prevalence across patient age brackets (left) and anatomical localisations (right) in HAM10000. The non-uniform risk distribution across these metadata groups provides empirical support for the tabular neurosymbolic risk score used by EcoFair to guide selective escalation.
• Figure 3: Pareto frontier on HAM10000 for the MobileNetV2--ResNet50 pair. Grey points denote non-optimal routing configurations obtained from threshold sweeps, while the highlighted frontier shows the non-dominated trade-off between total edge energy and worst-group TPR on malignant classes.
• Figure 4: Subgroup-sensitive fairness analysis for Pair I (MobileNetV2--ResNet50) across the three benchmark datasets. The left plot shows worst-group TPR on malignant classes, and the right plot shows the corresponding TPR gap across demographic subgroups.