PRIVACY-PRESERVING SPLIT AND COLLABORATIVE TRAINING FOR LARGE AND SMALL NEURAL ARCHITECTURES: THREATS, PROTOCOLS, AND A UNIFIED FRAMEWORK

Abstract

This article presents a unified, publication-ready exposition of privacy-preserving collaborative learning architectures with emphasis on split-learning for one-dimensional convolutional neural networks (1D-CNNs) and split/fine-tuning paradigms applied to large language models (LLMs). Building on empirical and theoretical work in split learning (Abuadbba et al., 2020), secret-sharing primitives and randomization techniques from multiparty computation (Beaver, 1992; Chase et al., 2020), and recent findings of vulnerabilities in split-based LLM fine-tuning (Chen et al., 2024), we synthesize secure protocol building blocks and evaluate their applicability across multiple deployment scenarios including mobile LLM-enhanced applications (Chandra, 2025) and decentralized health monitoring (Sharma et al., 2024). The article details threat models, attacker capabilities, and formal desiderata for privacy-preserving collaborative training; describes how classic cryptographic primitives can be combined with split learning to mitigate identified leakage; and proposes a practical hybrid protocol that trades off computation, communication, and privacy. We also present a careful, text-based methodological narrative on measuring privacy leakage without resorting to equations, illustrate expected empirical behaviours through descriptive analysis, and discuss operational limitations and future research directions. The article aims to serve researchers and practitioners seeking to design, evaluate, and deploy collaborative learning systems that are resilient to the kinds of bidirectional attacks recently demonstrated against split fine-tuning.