[Paper Review] Augmented Rotation-Based Transformation for Privacy-Preserving Data Clustering

Dowon Hong, Aziz Mohaisen|arXiv (Cornell University)|Jun 10, 2010

Blind Source Separation Techniques11 references3 citations

TL;DR

This paper proposes Augmented Rotation-Based Transformation (ARBT), an enhanced privacy-preserving data clustering method that extends Multiple Rotation-Based Transformation (MRBT) to preserve data utility for conventional clustering while resisting apriori-knowledge Independent Component Analysis (AK-ICA) attacks. ARBT leverages the linearity of transformation to enable clustering on transformed data subsets, achieving both strong privacy and practical usability, as validated through a custom toolkit and empirical evaluation on computational overhead and privacy metrics.

ABSTRACT

Multiple rotation-based transformation (MRBT) was introduced recently for mitigating the apriori-knowledge independent component analysis (AK-ICA) attack on rotation-based transformation (RBT), which is used for privacy-preserving data clustering. MRBT is shown to mitigate the AK-ICA attack but at the expense of data utility by not enabling conventional clustering. In this paper, we extend the MRBT scheme and introduce an augmented rotation-based transformation (ARBT) scheme that utilizes linearity of transformation and that both mitigates the AK-ICA attack and enables conventional clustering on data subsets transformed using the MRBT. In order to demonstrate the computational feasibility aspect of ARBT along with RBT and MRBT, we develop a toolkit and use it to empirically compare the different schemes of privacy-preserving data clustering based on data transformation in terms of their overhead and privacy.

Motivation & Objective

• To address the limitation of MRBT, which mitigates AK-ICA attacks but sacrifices data utility by disabling conventional clustering.
• To develop a transformation scheme that preserves data utility for clustering while maintaining resistance to AK-ICA attacks.
• To ensure computational feasibility of the proposed scheme through implementation and empirical evaluation.
• To demonstrate that ARBT enables clustering on transformed data subsets without compromising privacy.

Proposed method

• ARBT extends MRBT by exploiting the linearity of transformation to allow clustering on individual transformed data subsets.
• The method applies multiple rotation matrices to data, with each subset transformed using a distinct rotation matrix.
• A shared transformation framework ensures that the overall structure remains protected from AK-ICA attacks.
• The scheme allows conventional clustering algorithms to be applied directly on each transformed subset, preserving utility.
• A custom toolkit was developed to implement and benchmark RBT, MRBT, and ARBT for comparative analysis.
• Performance evaluation focused on computational overhead and privacy strength using metrics derived from the attack model.

Experimental results

Research questions

• RQ1 Can a transformation scheme be designed that resists AK-ICA attacks while preserving data utility for conventional clustering?
• RQ2 How does the computational overhead of ARBT compare to RBT and MRBT in practical deployment?
• RQ3 To what extent does ARBT maintain privacy while enabling clustering on transformed data subsets?
• RQ4 Can the linearity of transformation be leveraged to allow clustering on individual transformed data segments without exposing sensitive patterns?
• RQ5 What is the empirical trade-off between privacy protection and data utility in ARBT compared to existing schemes?

Key findings

• ARBT successfully resists AK-ICA attacks, maintaining strong privacy guarantees similar to MRBT.
• Unlike MRBT, ARBT enables conventional clustering on transformed data subsets, preserving data utility.
• The computational overhead of ARBT is comparable to RBT and MRBT, demonstrating practical feasibility.
• Empirical evaluation using the developed toolkit confirms that ARBT achieves a favorable balance between privacy and utility.
• The toolkit results show that ARBT outperforms MRBT in utility while maintaining equivalent or better privacy protection.
• The study confirms that linearity of transformation is a key enabler for enabling clustering on transformed data without compromising security.