Get Your Foes Fooled: Proximal Gradient Split Learning for Defense Against Model Inversion Attacks on IoMT Data

Sunder Ali Khowaja, Ik Hyun Lee, Kapal Dev, Muhammad Aslam Jarwar, Nawab Muhammad Faseeh Qureshi

TL;DR

This paper introduces Proximal Gradient Split Learning (PGSL), a novel defense mechanism against model inversion attacks on Internet of Medical Things (IoMT) data. PGSL intentionally introduces subtle pixel-level attacks to IoMT data during training, making it harder for attackers to reconstruct sensitive information from model gradients, while simultaneously improving recognition accuracy.

What Problem Does This Paper Address?

The increasing integration of deep learning (DL) in the Internet of Medical Things (IoMT) ecosystem, while promising for healthcare advancements, introduces significant privacy risks. IoMT data, containing highly sensitive personal health information, is vulnerable to various attacks, particularly model inversion attacks. These attacks exploit intermediate outputs of DL models to reconstruct the original input data, compromising patient privacy and potentially leading to incorrect diagnoses. Existing defense mechanisms often degrade model performance or are ineffective against sophisticated attacks. This paper aims to bridge this gap by proposing a defense that not only protects against model inversion but also enhances recognition accuracy.

Methods

The core of the proposed solution is the Proximal Gradient Split Learning (PGSL) framework. Here's a breakdown of its key components:

  • Intentional Pixel Attack: At the client-side, before data is sent for processing, a subtle, non-targeted pixel attack (using the Jacobian-based saliency map attack - JSMA) is applied to the IoMT data. This makes the raw data slightly "corrupted" in a controlled manner.
  • Split Learning Architecture: The network is split, with the client handling the initial processing (including the pixel attack and subsampling) and the server handling the more computationally intensive tasks.
  • Reversible Subsampling and Upsampling: A layer is used to downsample the attacked image on the client and upsample it on the server. This helps in retaining depth while increasing the receptive field.
  • Proximal Gradient for Recovery: On the server-side, a modified proximal gradient method is employed. This method aims to recover the original, unattacked image from the attacked input and its corresponding saliency map. This is crucial because the attacker typically only has access to intermediate feature maps, not the original saliency map.
  • Fusion Strategies: Three streams of data are processed: the attacked image, the fused image (combining attacked and recovered data), and the recovered image. These streams are then combined using a weighted-averaging decision-level fusion strategy to derive the final classification label. This fusion aims to leverage the strengths of each stream and improve overall robustness.

What's novel here is the intentional introduction of a controlled attack as a defense mechanism, combined with the application of proximal gradients for data recovery in a split learning context, specifically targeting model inversion attacks.

Datasets

The paper evaluates the proposed PGSL method on two datasets:

  • MIAS (Mammographic Image Analysis Society) dataset: A publicly available dataset for mammograms, relevant to medical imaging.
  • MNIST dataset: A widely used benchmark for handwritten digit recognition. The MNIST dataset was primarily used for fairness of comparison with existing approaches and for visual clarity of results.

Key Insights and Ideas

  • Controlled Attack as Defense: The counter-intuitive idea of intentionally attacking data to defend it is a key insight. By introducing a controlled, subtle attack, the system becomes more resilient to external, malicious attacks.
  • Proximal Gradients for Privacy: The application of proximal gradient methods for recovering attacked data in a split learning scenario is a novel approach to privacy preservation. It allows for the reconstruction of data without necessarily revealing the original sensitive information to the attacker.
  • Fusion for Robustness: Combining information from multiple streams (attacked, fused, and recovered) through decision-level fusion significantly enhances the model's resilience and recognition performance. This multi-pronged approach makes it harder for an attacker to succeed.
  • IoMT Specificity: The work directly addresses the unique challenges of IoMT data, where privacy is paramount and the consequences of data breaches or misdiagnoses are severe.

Key Results

  • Improved Recognition Accuracy: PGSL achieved significant accuracy gains. For example, it reported 14.0%, 17.9%, and 36.9% gains in accuracy over reconstructed and adversarial attacked images, respectively.
  • Defense Against Model Inversion: The proposed method demonstrated effectiveness in reducing the accuracy of attacks, with PGSL achieving 83.2% accuracy on MIAS and 99.8% on MNIST against various attacks, outperforming several state-of-the-art methods.
  • Reduced Reconstruction Performance: The paper shows that PGSL helps in reducing the recognition performance from reconstructed images, thereby reducing the degree of recovery from gradient maps.
  • Quantitative Comparison: Table I and II provide detailed quantitative comparisons of PGSL against various attack and recovery methods, showcasing its superior performance in terms of Mean Squared Error (MSE) for reconstruction and accuracy for recognition.

Limitations and Future Work

  • Hyperparameter Tuning: The effectiveness of the proximal gradient method relies on a hyperparameter, $\mu$. While the paper analyzes its sensitivity and selects an optimal value, fine-tuning for different datasets and modalities is necessary.
  • Assumption of Map Availability: The PGSL method assumes that the saliency map is available from the data aggregator stage. The authors acknowledge that advancements in GANs might require evolving this defense mechanism.
  • Accuracy Improvement Room: While PGSL shows promising results, the authors note that there is still room for improvement in achieved accuracy, especially considering the sensitivity of IoMT data where slight perturbations can lead to wrong diagnoses.
  • Future Directions:
    • Integration with Private AI frameworks for enhanced data and model security.
  • Observing the effect of PGSL in virtual worlds like Metaverse and Spatial computing.
  • Extending the work to observe its effect on multiple client/server nodes with split learning concerning domain adaptation.
  • Exploring other IoMT modalities beyond mammograms, such as X-ray, fundus images, CT-scans, and electronic health records.

Summary updated

Abstract

The past decade has seen a rapid adoption of Artificial Intelligence (AI), specifically the deep learning networks, in Internet of Medical Things (IoMT) ecosystem. However, it has been shown recently that the deep learning networks can be exploited by adversarial attacks that not only make IoMT vulnerable to the data theft but also to the manipulation of medical diagnosis. The existing studies consider adding noise to the raw IoMT data or model parameters which not only reduces the overall performance concerning medical inferences but also is ineffective to the likes of deep leakage from gradients method. In this work, we propose proximal gradient split learning (PSGL) method for defense against the model inversion attacks. The proposed method intentionally attacks the IoMT data when undergoing the deep neural network training process at client side. We propose the use of proximal gradient method to recover gradient maps and a decision-level fusion strategy to improve the recognition performance. Extensive analysis show that the PGSL not only provides effective defense mechanism against the model inversion attacks but also helps in improving the recognition performance on publicly available datasets. We report 14.0$\%$, 17.9$\%$, and 36.9$\%$ gains in accuracy over reconstructed and adversarial attacked images, respectively.
BibTeX
@article{khowaja2023get,
  author = {Sunder Ali Khowaja and Ik Hyun Lee and Kapal Dev and Muhammad Aslam Jarwar and Nawab Muhammad Faseeh Qureshi},
  title = {{Get Your Foes Fooled: Proximal Gradient Split Learning for Defense Against Model Inversion Attacks on IoMT Data}},
  journal = {IEEE Trans. Network Science \& Engineering},
  year = {2023},
  doi = {10.1109/tnse.2022.3188575},
}
Figure