Explainability-preserving, Privacy-aware Multi-teacher Distillation for Regulated Enterprise AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Knowledge distillation is widely used to compress large models into smaller student models that are easier to deploy. In regulated enterprise environments, however, compression alone is insufficient. Deployed systems may also need explanation support, bounded information transfer, documentation, traceability, and human oversight, especially in finance, healthcare, insurance, and public administration.This paper proposes a conceptual framework for explainability-preserving, privacy-aware multi-teacher distillation for regulated enterprise AI. The framework integrates four design layers: a multi-teacher distillation layer, an explainability-preservation layer, a privacy-aware training layer, and a governance and documentation layer. Rather than treating predictive transfer as the sole objective, the framework defines distillation as a constrained and auditable knowledge-transfer process in which a compact student model is designed to preserve task performance while also supporting selected explanation behavior, restricted transfer of sensitive information, and operational accountability.The paper makes four contributions. First, it synthesizes literature on knowledge distillation, multi-teacher learning, explainable AI, privacy-preserving machine learning, and AI governance into a unified design perspective for regulated deployment. Second, it identifies key design requirements for multi-objective distillation, including explanation alignment, teacher aggregation, privacy threat modeling, and traceability. Third, it provides not only a modular architecture and a reproducible objective template, but also one bounded illustrative instantiation for a credit-risk decision-support example covering aggregation, explanation, privacy, and governance choices. Fourth, it outlines an evaluation structure for future empirical validation in regulated use cases.The paper does not claim empirical performance gains, formal privacy guarantees for all deployments, or legal compliance. Instead, it provides a structured conceptual foundation and one internally specified illustrative profile intended to make future technical assessment more concrete.