Abstract
Running deep neural networks for large medical images is a resource-hungry and time-consuming task with centralized computing. Outsourcing such medical image processing tasks to hybrid clouds has benefits, such as a significant reduction of execution time and monetary cost. However, due to privacy concerns, it is still challenging to process sensitive medical images over clouds, which would hinder their deployment in many real-world applications. To overcome this, we first formulate the overall optimization objectives of the privacy-preserving distributed system model, i.e., minimizing the amount of information about the private data learned by the adversaries throughout the process, reducing the maximum execution time and cost under the user budget constraint. We propose a novel privacy-preserving and cost-effective method called PriCE to solve this multi-objective optimization problem. We performed extensive simulation experiments for artifact detection tasks on medical images using an ensemble of five deep convolutional neural network inferences as the workflow task. Experimental results show that PriCE successfully splits a wide range of input gigapixel medical images with graph-coloring-based strategies, yielding desired output utility and lowering the privacy risk, makespan, and monetary cost under user's budget.
Introduction: The Privacy-Performance Paradox
Processing gigapixel medical images, such as Whole Slide Images (WSIs) used in pathology, presents a massive computational challenge. A single image can contain billions of pixels, making deep learning inference prohibitively slow on local hardware. While Hybrid Clouds (HCs) offer a solution by distributing the workload, they introduce a critical trade-off between privacy and performance. Traditional cloud scheduling models often treat privacy as a binary constraint—either data is sent to the cloud or it isn't.
As shown in the comparison above, most existing research focuses on cost and time, treating privacy as a fixed constraint. The PriCE (Privacy-Preserving and Cost-Effective) framework shifts this paradigm by treating privacy as a quantifiable optimization objective. This allows for a more nuanced approach where the system simultaneously minimizes privacy risk, monetary cost, and total execution time (makespan).
The PriCE Framework
The PriCE method operates through a sophisticated six-step pipeline designed to manage the complexities of medical image processing in a distributed environment.
1. Graph-Coloring Image Splitting
To prevent an adversary from reconstructing a sensitive image, PriCE ensures that adjacent image patches—which share the most structural information—are never sent to the same cloud node. By modeling the image as a grid graph and applying graph-coloring algorithms (such as largest_first or independent_set), the system effectively "scrambles" the image across the hybrid cloud. No single compromised node possesses enough context to reconstruct the original data.
2. Workload Transformation and Coordinate Perturbation
The CNN inference workload is transformed into a series of tasks. To further protect data, PriCE uses coordinate perturbation. By applying techniques similar to Principal Component Analysis (PCA), the system transforms and normalizes coordinate data, adding random noise to prevent reverse-engineering of the original image structure while maintaining the ability to reconstruct the final results.
3. Multi-Objective Optimization (MOO)
The core of PriCE is its MOO problem-solving space. It quantifies privacy risk using Mutual Information, measuring how much an adversary can learn about the private data by observing the processed data. Because it is impossible to maximize privacy, speed, and cost-savings simultaneously, PriCE calculates a 3D Pareto Front. This set of optimal solutions allows users to select a configuration that best fits their specific budget and risk tolerance.
Experimental Results and Insights
The study utilized real-world medical data from The Cancer Genome Atlas (TCGA), specifically focusing on a Whole Slide Image (WSI) for artifact detection using an ensemble of five deep CNNs. The infrastructure was a simulated hybrid cloud consisting of 25 commercial GPU servers and 2 private GPU servers.
Key Findings: * Perfect Utility: Despite the complex encryption and coordinate perturbation, the system achieved perfect output utility, meaning the final reconstructed results were identical to those processed locally. * Privacy Gains: Graph-coloring strategies consistently outperformed standard random splitting methods, resulting in significantly lower average minimal privacy risk scores. * The Budget Effect: Research indicated that as budgets tighten, privacy is often the first objective to be compromised. However, the graph-coloring strategies in PriCE showed greater resilience in maintaining privacy under tight financial constraints compared to traditional methods.
Conclusion
PriCE demonstrates that privacy in medical AI does not have to be a binary "yes or no" choice. By treating privacy as a mathematical variable within a multi-objective optimization framework, researchers can deploy resource-hungry deep learning models on hybrid clouds without sacrificing data security. Future work will look toward addressing active malicious threat models and further optimizing the communication overhead inherent in secure distributed messaging.
Summary updated