Trust Aware Federated Learning for Secure Bone Healing Stage Interpretation in e-Health

This paper proposes a trust-aware federated learning framework that utilizes an Adaptive Trust Score Scaling and Filtering mechanism to secure bone healing stage interpretation in e-Health by mitigating the impact of unreliable or adversarial participants while maintaining model integrity and predictive performance.

The Gist

Imagine you are trying to teach a robot how to tell the difference between a fresh broken bone and a fully healed one. Instead of gathering all the X-rays and sensor data from every hospital into one giant, risky database (which would violate patient privacy), you want the hospitals to teach the robot together, keeping their data on their own computers. This is called Federated Learning.

However, there's a problem: What if one hospital has a broken sensor? What if another hospital's computer is glitching? Or worse, what if a "bad actor" tries to trick the robot with fake data? If the robot listens to everyone equally, the bad data will ruin its learning, and it might think a broken bone is healed (or vice versa).

This paper presents a smart solution called Trust-Aware Federated Learning. Here is how it works, explained with simple analogies:

1. The Problem: The "Noisy Classroom"

Imagine a classroom where 100 students are trying to solve a puzzle together. The teacher (the central server) asks everyone to share their answer.

• The Good Students: Most students are working hard and have the right answer.
• The Glitchy Students: Some students have bad eyesight or tired brains, so they give wrong answers by accident.
• The Pranksters: A few students might be trying to sabotage the puzzle on purpose.

If the teacher just averages everyone's answers (the standard method), the pranksters and the glitchy students will drag the whole class's grade down.

2. The Solution: The "Trust Score" System

The authors created a system that acts like a strict but fair referee. Instead of listening to everyone equally, the referee gives every student a "Trust Score."

• How the Score is Calculated: Every time a student shares an answer, the referee checks: "Is your answer accurate? Is it consistent with what you said before?" They use a complex math tool (called TOPSIS) to combine these checks into a single number between 0 and 1.
• The "Exclusion Zone": If a student's trust score drops below a certain line (like 0.75), the referee says, "Sorry, you're not allowed to speak this round." This stops the bad data from ruining the group's progress.
• The "Second Chance": If a student who was excluded starts giving good answers again, their score goes up, and they are let back into the game. This ensures the system doesn't permanently punish someone who just had a bad day.

3. The Secret Sauce: "Adaptive Smoothing"

This is the most clever part of the paper. Usually, referees might be too strict or too slow to react. The authors added a feature called Adaptive EMA (Exponential Moving Average).

Think of this like a thermostat or a shock absorber in a car:

• Static Mode (Old Way): The referee uses a fixed rule. If a student's score wobbles a little, the referee might overreact and kick them out, or ignore a real problem.
• Adaptive Mode (New Way): The referee watches how much the students' scores are shaking.
  • If everyone is calm and scores are steady, the referee reacts quickly to changes.
  • If everyone is chaotic and scores are jumping around wildly, the referee slows down. It says, "Okay, things are messy right now; I'll wait a bit to see if this is a real problem or just noise before I make a decision."

This "shock absorber" prevents the system from panicking when a student has a temporary glitch, keeping the learning process smooth and stable.

4. The Result: A Smarter Robot

The researchers tested this on simulated data representing bone healing stages (from a fresh break to a fully healed bone).

• Without the Trust System: The robot got about 67% of the bone stages right. It got confused easily when the data was messy.
• With the Trust System (Adaptive): The robot got 77.6% right. It became much better at telling the difference between tricky, similar-looking stages (like "early healing" vs. "almost healed").

Why This Matters for Real Life

In the real world, hospitals can't share patient data due to privacy laws. This system allows hospitals to collaborate safely. Even if one hospital has a broken sensor or a hacker tries to interfere, the "Trust Referee" filters them out, ensuring the medical AI learns correctly without ever seeing the private patient data.

In short: This paper teaches us how to build a team of AI doctors that can work together securely, ignore the "bad apples," and stay calm even when the data gets messy, leading to better diagnoses for broken bones.

Technical Summary

Here is a detailed technical summary of the paper "Trust Aware Federated Learning for Secure Bone Healing Stage Interpretation in e-Health."

1. Problem Statement

The paper addresses two primary challenges in applying Federated Learning (FL) to healthcare, specifically for bone healing stage classification:

• Data Privacy and Silos: Medical institutions cannot centralize patient data due to privacy regulations (e.g., GDPR, HIPAA). FL allows collaborative model training without sharing raw data, but standard FL algorithms struggle with data heterogeneity and unreliable participants.
• Adversarial and Unreliable Clients: In distributed medical sensing environments, clients (hospitals or devices) may provide noisy, corrupted, or adversarial updates. Standard aggregation methods like Federated Averaging (FedAvg) treat all clients equally, making the global model vulnerable to performance degradation caused by low-quality or malicious contributions.
• Specific Application Gap: While FL is used for medical imaging, its application to orthopedic frequency response data (spectral features for bone healing) is under-explored. Existing robust aggregation methods often rely on fixed rules or predefined attack assumptions, lacking the flexibility to adapt to dynamic, real-world participation patterns.

2. Methodology

The authors propose a Trust-Aware Federated Learning Framework that integrates the Adaptive Trust Score Scaling and Filtering (ATSSSF) mechanism with Exponential Moving Average (EMA) smoothing.

A. Data and Model Setup

• Dataset: Spectral measurements (S-parameters $S_{11}$ and $S_{21}$) across a frequency sweep, representing seven bone healing stages (from "Fresh Fracture" to "Fully Healed").
• Preprocessing: Features include spectral differences, ratios, and derivatives. Data was normalized, and class imbalance was addressed using SMOTE (Synthetic Minority Over-sampling Technique) on the training set.
• Model: A feedforward Multi-Layer Perceptron (MLP) with three dense layers, ReLU activation, dropout, and batch normalization, outputting 7 classes via Softmax.
• Simulation: The Flower framework was used to simulate 100 virtual clients over 500 communication rounds. Client heterogeneity and unreliability were synthetically introduced to test robustness.

B. The ATSSSF Mechanism

The core innovation is a dynamic client evaluation process:

1. Multi-Criteria Trust Scoring (TOPSIS): Instead of relying on a single metric, client contributions are evaluated using four local performance indicators: Accuracy, Precision, Recall, and F1-score. These are aggregated using the TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) method to generate a normalized trust score ($T_i \in [0, 1]$).
2. Trust Smoothing (EMA): To prevent transient fluctuations from causing erratic client exclusion, trust scores are smoothed using an Exponential Moving Average.
   • Static EMA: Uses a fixed smoothing coefficient ($\alpha = 0.3$).
   • Adaptive EMA: The smoothing coefficient $\alpha$ adjusts dynamically based on the variance of trust scores. High variance reduces $\alpha$ (emphasizing stability), while low variance increases $\alpha$ (allowing faster response to genuine improvements).
3. Filtering and Readmission:
   • Clients with a smoothed trust score below a threshold ($\tau = 0.75$) are omitted from the global aggregation.
   • A conservative cap limits omissions to a maximum of 3 clients per round to ensure aggregation continuity.
   • Readmission: Clients are re-admitted if their smoothed trust score exceeds the threshold for two consecutive rounds, allowing recovery from temporary instability.

3. Key Contributions

• Novel Application: First application of trust-aware FL to bone healing stage prediction using spectral frequency response data, a domain previously unexplored in FL literature.
• Adaptive Trust Management: Introduction of an Adaptive EMA strategy that dynamically adjusts smoothing based on trust score variance, offering a balance between stability (filtering noise) and responsiveness (adapting to real performance shifts).
• Multi-Metric Evaluation: Moving beyond simple accuracy, the framework uses TOPSIS to synthesize multiple performance metrics (Precision, Recall, F1) for a holistic trust assessment.
• Dynamic Client Lifecycle: A mechanism for omission and readmission that maintains model integrity without permanently excluding potentially recoverable clients, ensuring inclusivity in the federated network.

4. Experimental Results

The study compared the proposed ATSSSF framework (with static and adaptive EMA) against a baseline FedAvg model.

• Performance Metrics:
  • Baseline FedAvg: 67.4% Accuracy, 0.61 Macro F1.
  • ATSSSF (Static EMA): 75.1% Accuracy, 0.72 Macro F1.
  • ATSSSF (Adaptive EMA): 77.6% Accuracy, 0.74 Macro F1, and 0.76 Precision.
• Key Findings:
  • The Adaptive EMA configuration achieved the best results, improving accuracy by ~10% over the baseline.
  • Stability: Trust score variance decreased from 0.12 (FedAvg) to 0.07 (Adaptive ATSSSF), indicating a more stable aggregation process.
  • Classification Improvement: Misclassifications between spectrally similar stages (e.g., "Soft Callus" and "Early Mineralization") were reduced by approximately 12% compared to the baseline.
  • Client Management: On average, only 2 clients were excluded per round, and the mean trust score converged to 0.81, demonstrating effective filtering without excessive data loss.

5. Significance and Future Work

• Significance: The paper establishes the feasibility of adaptive trust mechanisms in federated medical sensing. It demonstrates that dynamic trust management can significantly improve model stability and predictive performance in environments with unreliable or adversarial participants, without requiring centralized data.
• Limitations: The study relies on a controlled, synthetically partitioned dataset rather than real-world multi-patient clinical data. Client behaviors were simulated, which limits immediate clinical generalizability.
• Future Directions:
  • Deployment in real-world clinical settings with live sensor data.
  • Integration with Multi-Access Edge Computing (MEC) to assess latency and fault tolerance.
  • Comparative analysis against other robust algorithms (e.g., Krum, FLCert).
  • Development of fully adaptive thresholds for the ATSSSF framework.

In conclusion, this work provides a robust, privacy-preserving framework for collaborative bone healing monitoring, proving that adaptive trust filtering is essential for the reliability of FL in critical healthcare applications.