💻 computer science

Securing IoMT with Federated Learning: A Hybrid Deep Learning and Ensemble-Based Intrusion Detection Framework

This paper proposes a privacy-preserving federated learning framework that integrates a hybrid Deep Neural Network and XGBoost model to effectively detect cyber threats in Internet of Medical Things environments while addressing challenges related to data privacy, non-IID data distribution, and adversarial attacks.

Original authors: Ifeanyi Nwokoro, Edgar Osaghae, Saheed Kayode, Tombari Sibe

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ifeanyi Nwokoro, Edgar Osaghae, Saheed Kayode, Tombari Sibe

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive hospital network where thousands of smart devices—like heart monitors, insulin pumps, and smart beds—are constantly talking to each other to keep patients safe. This is the Internet of Medical Things (IoMT). While this technology is amazing, it's also a giant target for cybercriminals trying to hack into the system, steal patient data, or shut down life-saving equipment.

The authors of this paper, a team of researchers from universities in the US and Nigeria, propose a new way to protect these devices. They call it a "Federated Learning" system. Here is how their solution works, explained simply:

The Problem: The "Centralized" Bottleneck

Traditionally, to catch hackers, hospitals would send all their data (patient records, device logs, network traffic) to one giant central computer. This computer would analyze everything to spot attacks.

The Analogy: Imagine every patient in a hospital having to bring their entire medical diary to a single security guard's desk to be read.
The Issue: This is a privacy nightmare (violating laws like HIPAA and GDPR) and a logistical nightmare. It's too slow, and if the central computer gets hacked, all the data is gone.

The Solution: The "Secret Recipe" Approach

The researchers suggest a smarter way called Federated Learning. Instead of sending the data to a central boss, they send the "brain" (the AI model) to the local devices.

The Analogy: Imagine a master chef (the central server) who wants to create the perfect security recipe. Instead of asking every local restaurant to send their secret ingredients (patient data) to the chef, the chef sends the recipe book to each restaurant. The local chefs cook a small batch of the dish using their own local ingredients, taste it, and send back only the notes on how to improve the recipe. The chef never sees the ingredients, only the notes.
The Result: The global security model gets smarter, but no patient's private data ever leaves their local device.

The "Hybrid" Engine: Two Brains Working Together

The paper claims that using just one type of AI isn't enough. They built a Hybrid Framework that combines two different types of "brains":

Deep Neural Network (DNN): Think of this as a detective who is great at spotting complex, hidden patterns in a messy crime scene (like deep learning).
XGBoost: Think of this as a sharp, rule-based analyst who is excellent at making quick, accurate decisions based on specific facts (like a tree-based decision maker).

The Analogy: It's like hiring a team where one person is an expert at reading body language (DNN) and the other is an expert at analyzing financial ledgers (XGBoost). Together, they catch more criminals than either could alone.

The Security Shield: Protecting the "Notes"

Even though the devices only send "notes" (model updates) and not data, a clever hacker might still try to trick the system or steal information from those notes. The paper adds three layers of protection:

Differential Privacy: This adds a tiny bit of "static" or "noise" to the notes sent back to the chef.
- Analogy: It's like whispering the recipe improvements into a room full of people so that even if someone eavesdrops, they can't hear the exact words, but the chef can still understand the general idea.
Secure Aggregation: This ensures the chef can only see the combined result of all the notes, not what any single restaurant sent.
Robust Aggregation: If a hacker tries to send a fake note to ruin the recipe (a "poisoning attack"), the system uses a voting mechanism to ignore the bad notes and stick with the majority.

What Did They Find? (The Results)

The researchers tested their system on three different sets of data (simulating hospital networks and general IoT devices). Here is what happened:

Accuracy: Their system caught about 91.4% of attacks. This is almost as good as the "old way" (centralized systems) which caught about 90-99%, but without the privacy risks.
Speed: The system was incredibly fast, taking only 0.0014 seconds to test for an attack. This means it can work in real-time on small, battery-powered devices.
Privacy Cost: Adding the "noise" for privacy (Differential Privacy) only hurt the accuracy by less than 1%. This is a tiny price to pay for keeping patient data safe.
Handling Messy Data: Real-world hospitals have different types of devices with different data. The system handled this "messiness" (called non-IID data) very well, proving it works even when devices aren't all the same.

The Bottom Line

The paper concludes that you don't have to choose between security and privacy. By using this "Federated" approach, hospitals can build a super-smart, hybrid AI system that detects hackers effectively while ensuring that sensitive patient data never leaves the local device. It's a scalable, efficient, and safe way to protect the future of medical technology.

Technical Summary: Securing IoMT with Federated Learning

Problem Statement
The rapid proliferation of Internet of Medical Things (IoMT) devices has expanded healthcare connectivity but simultaneously increased exposure to sophisticated cyber threats, including distributed denial-of-service (DDoS) attacks, malware propagation, and data exfiltration. Traditional centralized Intrusion Detection Systems (IDS) are increasingly unsuitable for IoMT environments due to strict privacy regulations (e.g., HIPAA, GDPR), decentralized data ownership, device heterogeneity, and communication constraints. Centralized approaches require aggregating sensitive patient data at a single server, creating significant privacy risks and regulatory hurdles. Furthermore, existing Federated Learning (FL) based IDS solutions often struggle with non-independent and identically distributed (non-IID) data, lack robust defenses against adversarial poisoning, and rarely explore hybrid architectures that combine deep feature extraction with ensemble learning.

Methodology
The paper proposes a privacy-preserving federated intrusion detection framework that integrates a hybrid Deep Neural Network (DNN) and XGBoost model within a robust FL architecture. The methodology is structured as follows:

System Architecture: The framework operates in a federated setting with a central aggregation server and $K$ distributed IoMT clients (e.g., hospitals, medical devices). Clients retain local data and exchange only model updates, ensuring raw data locality.
Hybrid Model Design: The core detection mechanism combines a lightweight DNN for deep feature extraction with an XGBoost (gradient boosting) classifier. This hybrid approach leverages the complementary strengths of deep learning for pattern recognition and tree-based models for handling class imbalance and interpretability.
Federated Optimization: The system utilizes a weighted Federated Averaging (FedAvg) algorithm. To address the "straggler effect" and intermittent connectivity, the framework supports asynchronous updates and partial client participation (20–50% of clients per round).
Privacy and Security Enhancements:
- Differential Privacy (DP): Gaussian noise is added to local gradients to prevent model inversion and membership inference attacks, satisfying $(\epsilon, \delta)$ -differential privacy guarantees.
- Secure Aggregation: Cryptographic techniques ensure the server cannot reverse-engineer individual client parameters.
- Robust Aggregation: To mitigate malicious client attacks (e.g., model poisoning, label flipping), the framework employs Byzantine-resilient aggregation strategies, specifically Median and Krum algorithms, alongside client reputation scoring.
Data and Preprocessing: Experiments utilize three benchmark datasets: UNSW-NB15, BoT-IoT, and WUSTL-EHMS-2020. Data is partitioned across clients using non-IID strategies (label-skew and quantity-skew) to simulate realistic heterogeneous IoMT environments. Feature engineering involves a hybrid selection process combining correlation analysis, recursive feature elimination, and tree-based importance scores.
Mathematical Formulation: The problem is modeled as distributed empirical risk minimization. The paper provides formal convergence analysis under bounded gradient variance and client drift, quantifying the impact of non-IID data using Jensen–Shannon divergence. Communication complexity is also formally modeled to assess bandwidth scalability.

Key Contributions

Novel Hybrid Framework: Introduction of a federated IDS integrating lightweight DNN and XGBoost, designed specifically for the resource-constrained and heterogeneous nature of IoMT.
Enhanced Privacy and Robustness: A training mechanism that combines differential privacy and secure aggregation to protect sensitive medical data, coupled with robust aggregation strategies (Median/Krum) to defend against adversarial poisoning.
Formal Analysis: A theoretical convergence analysis under non-IID data distributions, accounting for client drift and bounded gradient variance, alongside a formal quantification of communication overhead.
Comprehensive Evaluation: Extensive experiments across multiple datasets demonstrating the framework's performance in terms of accuracy, robustness against stragglers and attacks, and communication efficiency.

Results
Extensive experiments on the WUSTL-EHMS-2020, UNSW-NB15, and BoT-IoT datasets yielded the following findings:

Detection Performance: The proposed FL framework achieved a detection accuracy of up to 91.4% with F1-scores exceeding 90%. This performance closely approaches centralized baselines such as XGBoost (99% accuracy) and LightGBM (93% accuracy), while significantly outperforming standalone Random Forest and LSTM models in the federated context.
Efficiency: The framework demonstrated high computational efficiency, with testing times as low as 0.0014 seconds, making it suitable for real-time IoMT edge deployment.
Privacy-Utility Trade-off: The introduction of differential privacy resulted in less than 1% performance degradation under realistic noise budgets, validating the balance between privacy and utility.
Robustness: The system maintained high performance (F1 > 90%) under non-IID data heterogeneity, straggler effects, and adversarial model poisoning.
Communication: The framework showed efficient communication overhead, with total training times remaining low even with 20 clients, and convergence typically achieved within 50–60 rounds.

Significance
The paper claims that the proposed approach establishes a scalable, privacy-compliant, and efficient solution for securing next-generation IoMT systems. By bridging the gap between centralized IDS methods and privacy-aware, resource-efficient federated approaches, the framework enables the deployment of trustworthy, AI-driven cybersecurity in sensitive medical environments. The study highlights that hybrid deep learning and ensemble architectures can effectively balance predictive accuracy, robustness, and computational efficiency in federated IoMT settings, addressing critical gaps left by existing literature regarding non-IID handling, formal convergence guarantees, and combined privacy-robustness mechanisms.