[Paper Review] Privacy-Preserving SVM Computing by Using Random Unitary Transformation

Takahiro Maekawa, Takayuki Nakachi+2 more|arXiv (Cornell University)|Sep 19, 2018

Advanced Steganography and Watermarking Techniques19 references4 citations

TL;DR

This paper proposes a privacy-preserving SVM computing scheme that protects feature templates using random unitary transformations, enabling secure cloud-based SVM classification without performance degradation. The method preserves classification accuracy under isotropic stationary kernels and supports standard SVM algorithms without specialized encryption logic, validated through face recognition experiments with low error rates even under key or template leakage.

ABSTRACT

A privacy-preserving Support Vector Machine (SVM) computing scheme is proposed in this paper. Cloud computing has been spreading in many fields. However, the cloud computing has some serious issues for end users, such as unauthorized use and leak of data, and privacy compromise. We focus on templates protected by using a random unitary transformation, and consider some properties of the protected templates for secure SVM computing, where templates mean features extracted from data. The proposed scheme enables us not only to protect templates, but also to have the same performance as that of unprotected templates under some useful kernel functions. Moreover, it can be directly carried out by using well-known SVM algorithms, without preparing any algorithms specialized for secure SVM computing. In the experiments, the proposed scheme is applied to a face-based authentication algorithm with SVM classifiers to confirm the effectiveness.

Motivation & Objective

• To address privacy risks in cloud-based SVM computing due to data leaks and unauthorized access.
• To develop a method that protects feature templates (biometric or otherwise) without degrading SVM classification performance.
• To enable secure SVM inference in untrusted environments using standard SVM algorithms, avoiding specialized secure computation frameworks.
• To enhance security against spoofing attacks through key-based access control and template protection.

Proposed method

• Protects feature templates via random unitary transformation using secret keys, ensuring confidentiality.
• Applies the transformation to input features before sending them to the cloud, preserving inner product relationships under specific kernel functions.
• Uses isotropic stationary kernels (e.g., RBF, linear) where the transformed templates maintain equivalent discriminative power.
• Employs two key conditions: shared key (all users use same key) and unique key (each user has a distinct key), to control access and security granularity.
• Leverages the kernel trick to perform SVM classification directly on protected templates without decryption.
• Uses down-sampling to generate 1216-dimensional templates from facial images, which are then protected via random permutation matrices.

Experimental results

Research questions

• RQ1 Can random unitary transformation preserve the discriminative power of SVM classifiers when applied to protected templates?
• RQ2 Does the proposed method maintain the same classification accuracy as unprotected templates under standard kernel functions?
• RQ3 How does the use of shared vs. unique keys affect security and performance in a multi-user setting?
• RQ4 To what extent does the scheme resist spoofing attacks when keys or original templates are leaked?
• RQ5 Can standard SVM algorithms be directly used on protected templates without algorithmic modification?

Key findings

• The proposed scheme achieves identical classification performance between protected and unprotected templates under key condition 1 (shared key), with Equal Error Rate (EER) values matching across both cases.
• Under key condition 2 (unique keys per user), the False Accept Rate (FAR) decreased significantly due to stricter authentication requirements, improving security.
• Even when a key was leaked, the FAR remained low (slightly degraded from condition 2), demonstrating resilience to spoofing attacks.
• When original templates were leaked, the FAR remained low, confirming that protection via unitary transformation effectively prevents unauthorized access.
• The method enables direct use of standard SVM algorithms on protected data, eliminating the need for specialized secure computation frameworks.
• Face recognition experiments using Extended Yale Face Database B confirmed high accuracy with RBF and linear kernels, validating the practicality of the approach.