[Paper Review] MPCLeague: Robust MPC Platform for Privacy-Preserving Machine Learning

Ajith Suresh|arXiv (Cornell University)|Dec 26, 2021

Stochastic Gradient Optimization Techniques4 citations

TL;DR

MPCLeague is a high-performance, robust multi-party computation platform for privacy-preserving machine learning (PPML) in the secure outsourced computation (SOC) model, designed for 2–4 parties with up to one corruption. It achieves orders-of-magnitude speedup over prior work by optimizing secure arithmetic circuits and leveraging efficient secret-sharing techniques, demonstrating superior throughput and cost efficiency in benchmarks for logistic regression, neural networks, and SVMs on Google Cloud.

ABSTRACT

In the modern era of computing, machine learning tools have demonstrated their potential in vital sectors, such as healthcare and finance, to derive proper inferences. The sensitive and confidential nature of the data in such sectors raises genuine concerns for data privacy. This motivated the area of Privacy-preserving Machine Learning (PPML), where privacy of data is guaranteed. In this thesis, we design an efficient platform, MPCLeague, for PPML in the Secure Outsourced Computation (SOC) setting using Secure Multi-party Computation (MPC) techniques. MPC, the holy-grail problem of secure distributed computing, enables a set of n mutually distrusting parties to perform joint computation on their private inputs in a way that no coalition of t parties can learn more information than the output (privacy) or affect the true output of the computation (correctness). While MPC, in general, has been a subject of extensive research, the area of MPC with a small number of parties has drawn popularity of late mainly due to its application to real-time scenarios, efficiency and simplicity. This thesis focuses on designing efficient MPC frameworks for 2, 3 and 4 parties, with at most one corruption and supports ring structures. At the heart of this thesis are four frameworks - ASTRA, SWIFT, Tetrad, ABY2.0 - catered to different settings. The practicality of our framework is argued through improvements in the benchmarking of widely used ML algorithms -- Linear Regression, Logistic Regression, Neural Networks, and Support Vector Machines. We propose two variants for each of our frameworks, with one variant aiming to minimise the execution time while the other focuses on the monetary cost. The concrete efficiency gains of our frameworks coupled with the stronger security guarantee of robustness make our platform an ideal choice for a real-time deployment of PPML techniques.

Motivation & Objective

• To design a practical, efficient MPC framework for privacy-preserving machine learning (PPML) in the secure outsourced computation (SOC) setting.
• To achieve robustness in 2, 3, and 4-party MPC with up to one corruption, ensuring honest parties always receive the output regardless of corrupt behavior.
• To optimize performance for real-world PPML workloads such as logistic regression, neural networks, and support vector machines under ring-based arithmetic.
• To provide a unified, scalable platform that outperforms existing solutions in runtime, communication, throughput, and monetary cost.
• To explore extensions to advanced PPML applications and adversarial models beyond honest-majority settings.

Proposed method

• Designing a unified MPC protocol stack optimized for 2, 3, and 4 parties using arithmetic secret sharing over rings.
• Implementing efficient building blocks such as comparison, multiplication, and linear algebra operations tailored for low-latency execution.
• Integrating robustness mechanisms to prevent denial-of-service attacks from corrupt parties, ensuring output delivery even under malicious behavior.
• Using a two-phase protocol model: offline precomputation and online execution, with optimized online phase performance.
• Leveraging hardware-optimized implementations on Google Cloud to benchmark real-world performance across diverse workloads.
• Adopting a modular architecture that supports extensibility to new PPML primitives like batch normalization and square root.

Experimental results

Research questions

• RQ1 How can MPC protocols be made robust and efficient for small-party settings (2–4 parties) with one corruption in the context of PPML?
• RQ2 What architectural and algorithmic optimizations enable orders-of-magnitude improvements in online throughput and cost efficiency?
• RQ3 Can a unified MPC framework support diverse PPML workloads—such as neural networks, logistic regression, and SVMs—while maintaining strong security guarantees?
• RQ4 How does the platform perform under varying data sizes, batch sizes, and feature dimensions in real cloud deployments?
• RQ5 What are the implications of extending MPCLeague to federated learning or dishonest-majority adversarial models?

Key findings

• MPCLeague achieved an online throughput of 1317.96 iterations per minute for neural network training (NN-1) with 1000 features and batch size 1000, significantly outperforming SecureML and ABY2.0.
• ABY2.0T demonstrated the best performance in comparison benchmarks, achieving 128 comparisons per minute with a cost of only $0.05 in the 128-depth layer.
• For a 1024-depth comparison layer, MPCLeague’s framework achieved a throughput of 59.27 iterations per minute with a cost of $6.03, showing strong scalability.
• The platform reduced monetary cost by up to 80% compared to prior systems, with ABY2.0T being the most cost-effective in all tested configurations.
• The robustness mechanism ensured that honest parties always received the output, even when corrupt parties attempted to disrupt computation.
• The implementation on Google Cloud confirmed that MPCLeague is practical for real-world deployment, with measurable improvements in runtime, communication, and cost across all metrics.