Security and Privacy Issues and Solutions in Federated Learning for Digital Healthcare

Hyejun Jeong, Tai-Myoung Chung

Federated Learning (FL) is often hailed as the "holy grail" for digital healthcare. It promises the ability to train powerful AI models on sensitive patient data without that data ever leaving the hospital’s local servers. However, as a new survey paper by Hyejun Jeong and Tai-Myoung Chung from Sungkyunkwan University points out, this decentralized approach trades one set of privacy problems for a whole new "attack surface."

TL;DR

This paper provides a comprehensive taxonomy of the security and privacy vulnerabilities inherent in Federated Learning within the healthcare sector. It categorizes diverse threats—from "poisoned" models to "free-riding" participants—and evaluates the current state of defense mechanisms, ultimately calling for more standardized and multi-modal research to make FL truly "healthcare-ready."

What Problem Does This Paper Address?

Traditional AI requires "centralization": gathering all patient records into one giant database. In the era of GDPR and HIPAA, this is a legal and ethical nightmare. Federated Learning solves this by bringing the model to the data rather than the data to the model.

The catch? Because FL relies on communication between a central server and many distributed clients (hospitals, clinics, or even wearable devices), it introduces new vulnerabilities. An attacker doesn't need to breach a central database anymore; they can corrupt the model by pretending to be a hospital or intercepting the updates sent over the network. This paper maps out these risks specifically for the high-stakes world of digital health.

Methods

As a survey and position paper, the authors developed a systematic taxonomy to organize the "chaos" of FL security research. They categorize the field into three pillars:

  1. Sources of Vulnerabilities: Identifying who or what is at risk (Compromisable Clients, Compromisable Servers, Outsiders, and the Distributed Nature of the system).
  2. Attacks and Threats: Categorizing how systems are compromised (Poisoning, Inference, Backdoors, etc.).
  3. Mitigation Techniques: Evaluating current defenses (Sanitization, Perturbation, Cryptography, and Personalization).

The authors specifically highlight the "Severity" of different threats (Table 1 in the paper), ranking Poisoning, Inference, and Backdoor attacks as the highest risks to healthcare systems.

Datasets

The paper itself does not introduce a new dataset. Instead, it critiques the current state of the field. The authors observe that most existing FL security research relies on standard image benchmarks like:

  • MNIST (handwritten digits)
  • CIFAR-10 (general objects)

The authors point out a significant gap: these datasets do not represent the complexity of real-world healthcare data, which includes Electronic Health Records (EHR), medical imaging, and tabular data.

Key Insights and Ideas

  • Model Poisoning > Data Poisoning: The authors note that "Model Poisoning" (where a client directly alters the model parameters) is far more effective and stealthy than "Data Poisoning" (simply changing the training labels). In FL, a single malicious client can achieve 100% targeted misclassification while remaining undetected.
  • The "Honest-but-Curious" Server: One of the biggest threats is the central aggregator itself. Even if the server follows the rules, it can use "Inference Attacks" to reconstruct private patient data just by looking at the mathematical updates sent by hospitals.
  • The Non-IID Defense Dilemma: Healthcare data is naturally "Non-IID" (Non-Independent and Identically Distributed)—meaning a dermatology clinic's data looks very different from a cardiology clinic's data. Most current anomaly detection systems struggle to tell the difference between a "malicious" update and a "unique but legitimate" healthcare update.
  • Backdoors are the "Silent Killers": Backdoor attacks allow a model to function perfectly on normal cases but fail (or misdiagnose) when a specific "trigger" is present. These are incredibly difficult to detect because they don't degrade the overall accuracy of the model.

Key Results

The paper synthesizes the "state of the art" in defenses:

  • Differential Privacy (DP): Effective at stopping data leakage but often degrades model accuracy—a dangerous trade-off in medicine.
  • Robust Aggregation: Techniques like Krum or Bulyan help filter out malicious updates, but recent research shows they can be bypassed by sophisticated attackers.
  • Homomorphic Encryption (HE): Allows the server to compute updates without seeing them, but it is computationally expensive (though newer methods like BatchCrypt are making it faster).

Limitations and Future Work

The authors identify several "unmet needs" that the research community must address:

  1. Standardization: Different hospitals use different data formats. We need a standardized preprocessing pipeline to ensure models don't deviate simply because of formatting issues.
  2. Multi-modal Data: Research needs to move beyond simple image classification and test defenses on complex, multi-type healthcare data (text + images + tables).
  3. Incentive Mechanisms: Why should a large hospital share its valuable data if the system is risky? We need "Incentive Mechanisms" that reward high-quality, honest participation.
  4. The Defense Trade-off: Most current defenses are "orthogonal"—they solve one problem but create another (e.g., encryption stops snooping but causes communication bottlenecks). Future work needs to find "United Defense" techniques that balance privacy, security, and performance.

Summary updated

Abstract

The advent of Federated Learning has enabled the creation of a high-performing model as if it had been trained on a considerable amount of data. A multitude of participants and a server cooperatively train a model without the need for data disclosure or collection. The healthcare industry, where security and privacy are paramount, can substantially benefit from this new learning paradigm, as data collection is no longer feasible due to stringent data policies. Nonetheless, unaddressed challenges and insufficient attack mitigation are hampering its adoption. Attack surfaces differ from traditional centralized learning in that the server and clients communicate between each round of training. In this paper, we thus present vulnerabilities, attacks, and defenses based on the widened attack surfaces, as well as suggest promising new research directions toward a more robust FL.
BibTeX
@inproceedings{jeong2022security,
  author = {Hyejun Jeong and Tai-Myoung Chung},
  title = {{Security and Privacy Issues and Solutions in Federated Learning for Digital Healthcare}},
  booktitle = {FDSE (CCIS Volume)},
  year = {2022},
  url = {https://openreview.net/forum?id=9sgWm2fMKeo},
}

Loading related content...

Figure