Suda: An Efficient and Secure Unbalanced Data Alignment Framework for Vertical Privacy-Preserving Machine Learning
No ratings
Presented at USENIX Security 2025 by
Secure data alignment, which securely aligns the data between parties, is the first and crucial step in vertical privacy-preserving machine learning (VPPML). Practical applications, e.g. advertising, require VPPML for personalized services. Meanwhile, the data held by parties in these applications are usually unbalanced. Existing secure unbalanced data alignment approaches typically rely on Cuckoo Hashing, which introduces redundant data outside the intersection, leading to significantly increasing communication size during secure training in VPPML. Though secure shuffle operations can trim these redundant data, these operations would incur huge communication overhead. As a result, these secure approaches should be optimized for efficiency in VPPML scenarios.