Multi-Version Infrastructure for Privacy-Preserving AI/ML Inference at Scale

Read full text Summarise with VIRAT
Authors: Jay Bankimchandra Desai
Cited by: 0
References: 5

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

As the number of regulatory regimes, multi-stakeholder data relationships, and compliance requirements grows, privacy becomes an increasing architectural concern for large-scale AI/ML systems for data inference. Inference pipelines that apply a single, globally cast restrictive data policy to every inference context incur a measurable decrease in model performance. To avoid degrading model performance through globally restrictive policies while also avoiding potential policy violations introduced by dynamically modifying data usage per request, our multi-version architecture explicitly maintains multiple versions of user and participant information at the feature and embedding levels. In conjunction, context-aware version selection mechanisms deterministically map the metadata describing an incoming request to the appropriate data usage policy at runtime. In turn, versioned feature vectors are generated from superset representations of available signals, with the appropriate version selected based on the incoming request context and its corresponding data usage policy. Model-specific embeddings are derived from their privacy-compliant feature vectors to ensure end-to-end compliance. Rule-based selection schemes, implemented as abstractions decoupled from inference execution code, allow rapid regulatory adaptation without requiring service redeployment. Continuous monitoring helps validate selection quality and detect performance regressions in production environments. The computational overhead introduced by generating and maintaining multiple feature and embedding versions can be reduced through centralized build-once orchestration, shared feature storage schemas, and hybrid offline–online embedding generation within internet-scale latency budgets. Beyond privacy, this architectural pattern generalizes to fairness-aware inference, multi-tenant data isolation, and auditable policy enforcement, enabling versioned features and embedding representations as a foundational primitive for developing trustworthy, policy-compliant AI/ML systems. DOI: https://doi.org/10.17762/ijisae.v14i1s.8129

Fields of Study & Subjects

Privacy-Preserving Technologies in Data Ethics and Social Impacts of AI Privacy Security and Data Protection

Publication details

Journal / source
International Journal of Intelligent Systems and Applications in Engineering
Publication date
26 March 2026
Language
EN