Federated Learning over Human-Body Communication for On-Body Edge Intelligence: A Survey, Taxonomy, and BODYFED-HBC Scheduling Vignette

Technical Summary: Federated Learning over Human-Body Communication for On-Body Edge Intelligence

Problem Statement
The paper addresses a critical disconnect between two established research domains: Federated Learning (FL) for wearables and Human-Body Communication (HBC). While FL is recognized for reducing raw-data centralization in physiological sensing, and HBC is known for localizing communication and reducing radio leakage, they are rarely integrated. Existing wearable FL literature typically abstracts the communication layer as generic wireless networking, ignoring the unique constraints of HBC. Conversely, HBC research often abstracts away learning traffic. This gap creates a suboptimal system where learning protocols may select clients at inappropriate times, waste energy on retransmissions due to posture-dependent link failures, or bias models toward sensors with "easier" links, all while failing to fully leverage the privacy benefits of keeping raw data on the body.

Methodology
The paper proposes a framework for Body-Channel-Aware Federated Learning, where client selection, update compression, aggregation, and fairness are co-designed with the physical state of the HBC channel.

1. System Model: The authors define an intra-WBAN (Wireless Body-Area Network) scenario where on-body sensors (clients) transmit compressed model updates to a body hub via HBC/BCC links. The channel state is dynamic, dependent on posture, motion, and body location, affecting path loss, packet error rates, and energy consumption.
2. Optimization Formulation: A constrained optimization problem is formulated to minimize expected model loss while adhering to energy budgets, memory limits, and latency constraints. The objective function balances learning utility, energy cost, time cost, fairness across body locations, and privacy risk proxies.
3. Scheduling Algorithm (BODYFED-HBC): A greedy scheduling algorithm is introduced. It selects clients based on a decomposed utility function that accounts for:
   • Update Value and Novelty: The informational contribution of the local data.
   • Channel Cost: Estimated energy and packet success probability derived from HBC channel models.
   • Fairness Debt: A mechanism to prevent the exclusion of sensors with difficult links (e.g., distal sensors) by tracking historical participation.
   • Covariance Penalties: Terms that discourage selecting clients whose links fail simultaneously (correlated failures) and reward update diversity.
4. Bias Correction: To counteract the tendency of channel-aware schedulers to favor reliable links (e.g., chest-to-hub) over difficult ones (e.g., ankle-to-hub), the paper proposes an inverse-propensity reliability-aware aggregation method. This adjusts aggregation weights based on the moving-average probability of a client being selected, ensuring distal sensors contribute to the global model despite lower link reliability.
5. Reproducible Vignette: Recognizing the lack of public, posture-labeled HBC packet traces, the authors propose a computational vignette. This involves combining public wearable datasets (PAMAP2, MHEALTH, WESAD) with empirical BCC signal-loss datasets (Ormanis et al.) to emulate HBC-constrained FL rounds in a Jupyter environment.

Key Contributions
The paper makes six specific contributions:

1. Scoping Survey: A comprehensive review of FL for wearables, HBC/BCC physical layers, WBAN edge intelligence, and IoT privacy.
2. Taxonomy: A classification of learning architectures distinguishing between intra-WBAN FL, body-hub FL, cross-user FL, clinical/vendor FL, and hybrid split/distillation variants.
3. Privacy Distinction: A precise delineation of privacy layers, distinguishing between HBC link privacy (physical confinement) and raw-data/model-update privacy (cryptographic/statistical protection).
4. Reference Architecture (BODYFED-HBC): A concrete architecture including a constrained optimization formulation and a body-channel-aware scheduling algorithm.
5. Computational Blueprint: A reproducible simulation vignette using public datasets to demonstrate how to test body-channel-aware scheduling without requiring custom hardware.
6. Research Agenda: Identification of missing datasets, trace formats, and metrics, specifically calling for public, packet-level, posture-labeled HBC traces aligned with wearable tasks.

Results and Evaluation
As a survey and reference architecture paper, it does not present measured hardware results. Instead, it provides:

• Illustrative Emulator Output: Deterministic curves from a toy emulator demonstrating that the proposed scheduler can achieve target accuracy with lower energy-to-target ratios compared to naive baselines, without excluding difficult body locations.
• Break-Even Analysis: A theoretical calculation comparing the energy cost of raw streaming versus FL. The paper notes that FL is only attractive when the energy cost of local training and update transmission is lower than continuous raw streaming, or when privacy/cross-user learning justifies the extra cost.
• Evaluation Checklist: A proposed set of baselines (e.g., raw streaming, random selection, channel-only selection) and metrics (e.g., worst-location F1, energy-to-target-F1) for future experimental work.

Significance and Claims
The paper modestly positions itself as a scoping review and research agenda rather than a full experimental systems validation. Its primary significance lies in synthesizing the fields of HBC and FL into a coherent "body-channel-aware" framework.

• Bridging the Gap: It provides the necessary abstractions (channel-loss distributions, packet-success models, energy-per-bit estimates) for computer scientists to work on HBC-aware learning without designing analog transceivers.
• Defining the Problem: It argues that the "missing research area" is not just applying FL to wearables, but specifically adapting FL protocols to the dynamic, posture-dependent, and energy-constrained nature of the human body as a communication channel.
• Actionable Path: It offers a practical, low-risk path for researchers: start with public datasets and empirical loss models to validate scheduling logic before attempting full hardware integration. The paper concludes that while a full experimental systems paper requires measured traces, the proposed survey, taxonomy, and vignette provide a credible foundation for the field to advance.