Lightweight Multi-Class Support Vector Machine-Based Medical Diagnosis System with Privacy Preservation

Sherif Abdelfattah, Mohamed Baza, Mohamed Mahmoud, Mostafa M. Fouda, Khalid Abualsaud, Elias Yaacoub, Maazen Alsabaan, Mohsen GuizaniSensors 2023 · OpenReview

TL;DR

This paper introduces a novel, lightweight system for multi-class medical diagnosis using Support Vector Machines (SVMs) that prioritizes patient data privacy and protects the intellectual property of the diagnostic models. By modifying an inner product encryption cryptosystem, the system enables cloud-based diagnosis without revealing sensitive patient information or the proprietary model details to the cloud provider, achieving this with greater efficiency than previous methods.

What Problem Does This Paper Address?

The increasing adoption of cloud-based machine learning for medical diagnosis offers significant benefits, such as improved accessibility and accuracy. However, a critical challenge arises from the need to protect sensitive patient health data and the intellectual property of the diagnostic models themselves. When these models are outsourced to third-party cloud servers, there's a risk of data breaches or model theft. Existing solutions often struggle with high computational and communication overhead, fail to fully conceal classification results, or require complex multi-server setups. This paper tackles the gap by proposing an efficient, single-cloud system that addresses these privacy and security concerns for multi-class SVM medical diagnosis.

Methods

The core innovation of this paper lies in the modification of an inner product encryption cryptosystem. The original cryptosystem, designed for a single sender and single receiver, was adapted to a single-sender (model owner) and multiple-receiver (patients) scenario. This modification allows the model owner to encrypt their SVM model parameters once, which can then be used by all patients for diagnosis.

The system operates in four phases:

  1. System Initialization: A Key Distribution Center (KDC) generates secret keys for hospitals (model owners) and patients (users).
  2. Model Encryption: The hospital encrypts its multi-class SVM model parameters using its secret key. These encrypted parameters, along with masking numbers for privacy, are sent to the cloud server.
  3. Medical Data Encryption: Patients encrypt their medical data (symptoms, vital signs) using their unique secret keys.
  4. Classification: The cloud server performs computations on the encrypted model parameters and encrypted patient data to produce a masked classification result. This masked result is sent back to the patient, who can then decrypt and unmask it to obtain the final diagnosis.

The paper specifically leverages a One-Versus-All (1VA) approach for multi-class classification, which is more efficient than the One-Versus-One (1V1) method as it requires fewer binary classifiers.

Datasets

The authors used a dermatology diseases dataset from [19] for training and evaluating their SVM diagnosis model. This dataset contains 366 samples across six skin diseases, with 34 features per sample. The dataset is described as being widely used and considered a benchmark in medical applications.

Key Insights and Ideas

  • Efficiency through Cryptosystem Modification: The key insight is that adapting an existing inner product encryption scheme for a multi-user scenario dramatically reduces the computational burden on the model owner, a significant improvement over single-encryptor/single-decryptor designs.
  • Single-Cloud Advantage: The proposed system's reliance on a single cloud server simplifies deployment and reduces costs compared to multi-server solutions, while still achieving robust privacy guarantees.
  • Masking for Privacy: The use of random numbers to mask classification results before they are sent to the patient is crucial for preventing the cloud server from inferring any diagnostic information, even if it could somehow partially decrypt the results.
  • Linear SVM Suitability: The paper highlights that for privacy-preserving cloud-based systems, linear SVMs are often more practical than non-linear ones due to their lower computational and communication overhead in the encrypted domain.

Key Results

  • Privacy and Security: The proposed scheme is demonstrated to be secure against various attacks, including those from external eavesdroppers, known-ciphertext, known-background, and known-plaintext models, without revealing patient data or model parameters.
  • Computational Efficiency: The system significantly reduces computation costs for model owners and users compared to existing methods like those in [12] and [13]. For instance, the computation cost on model owners is shown to be constant regardless of the number of classes.
  • Communication Efficiency: The communication overhead is also minimized, particularly for the cloud server, which handles fewer complex operations.
  • Accuracy Maintenance: The privacy-preserving implementation using encrypted data achieves the same classification accuracy (97% for Accuracy, Precision, and F1-score) as the plaintext implementation, indicating no degradation in diagnostic performance.

Limitations and Future Work

  • Dataset Specificity: While the dermatology dataset serves as a good case study, the authors acknowledge that model parameters might change with different datasets, though their cryptosystem is designed to handle this.
  • Non-linear SVM Complexity: The paper notes that while non-linear SVMs can offer more sophisticated decision boundaries, their implementation in the ciphertext domain for privacy preservation is significantly more complex and computationally intensive, which is why they opted for linear SVMs.
  • Future Work: The paper suggests that the proposed methodology can be extended to other applications requiring privacy-preserving machine learning model evaluation, such as electricity theft detection, electric vehicle charging coordination, and vehicular social networks.

Summary updated

Abstract

Machine learning, powered by cloud servers, has found application in medical diagnosis, enhancing the capabilities of smart healthcare services. Research literature demonstrates that the support vector machine (SVM) consistently demonstrates remarkable accuracy in medical diagnosis. Nonetheless, safeguarding patients’ health data privacy and preserving the intellectual property of diagnosis models is of paramount importance. This concern arises from the common practice of outsourcing these models to third-party cloud servers that may not be entirely trustworthy. Few studies in the literature have delved into addressing these issues within SVM-based diagnosis systems. These studies, however, typically demand substantial communication and computational resources and may fail to conceal classification results and protect model intellectual property. This paper aims to tackle these limitations within a multi-class SVM medical diagnosis system. To achieve this, we have introduced modifications to an inner product encryption cryptosystem and incorporated it into our medical diagnosis framework. Notably, our cryptosystem proves to be more efficient than the Paillier and multi-party computation cryptography methods employed in previous research. Although we focus on a medical application in this paper, our approach can also be used for other applications that need the evaluation of machine learning models in a privacy-preserving way such as electricity theft detection in the smart grid, electric vehicle charging coordination, and vehicular social networks. To assess the performance and security of our approach, we conducted comprehensive analyses and experiments. Our findings demonstrate that our proposed method successfully fulfills our security and privacy objectives while maintaining high classification accuracy and minimizing communication and computational overhead.
BibTeX
@inproceedings{abdelfattah2023lightweight,
  author = {Sherif Abdelfattah and Mohamed Baza and Mohamed Mahmoud and Mostafa M. Fouda and Khalid Abualsaud and Elias Yaacoub and Maazen Alsabaan and Mohsen Guizani},
  title = {{Lightweight Multi-Class Support Vector Machine-Based Medical Diagnosis System with Privacy Preservation}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=iYXUaqUZTE},
}
Figure