NoEsis: Differentially Private Knowledge Transfer in Modular LLM Adaptation

2025-04-25

AI Summary

To address the challenge of simultaneously ensuring privacy preservation and enabling knowledge sharing in modular large language models (e.g., Mixture-of-Experts, MoE) during cross-domain adaptation, this paper proposes the first differentially private (DP) framework for two-stage parameter-efficient fine-tuning. In Stage I, a shared prompt backbone is trained via DP-SGD; in Stage II, private knowledge transfer is achieved using low-rank expert adapters. The method integrates differential privacy, LoRA, MoE, and prompt tuning. Evaluated on CodeXGLUE code completion, it provides provable privacy guarantees (ε ≤ 8), robustly resists membership inference attacks, and recovers over 77% of the accuracy gap between non-private/non-shared baselines. This significantly enhances the generalization capability and practical applicability of modular LLMs in privacy-sensitive domains.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Computer Vision: Diffusion Models for Vision

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Security and Privacy: Privacy-enhancing technologies
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large Language Models (LLM) are typically trained on vast amounts of data from various sources. Even when designed modularly (e.g., Mixture-of-Experts), LLMs can leak privacy on their sources. Conversely, training such models in isolation arguably prohibits generalization. To this end, we propose a framework, NoEsis, which builds upon the desired properties of modularity, privacy, and knowledge transfer. NoEsis integrates differential privacy with a hybrid two-staged parameter-efficient fine-tuning that combines domain-specific low-rank adapters, acting as experts, with common prompt tokens, acting as a knowledge-sharing backbone. Results from our evaluation on CodeXGLUE showcase that NoEsis can achieve provable privacy guarantees with tangible knowledge transfer across domains, and empirically show protection against Membership Inference Attacks. Finally, on code completion tasks, NoEsis bridges at least 77% of the accuracy gap between the non-shared and the non-private baseline.

Problem

• Ensuring privacy in modular LLM knowledge transfer
• Balancing generalization with isolated training limitations
• Achieving accuracy with differential privacy guarantees

Innovation

• Differential privacy for modular LLM adaptation
• Hybrid two-staged parameter-efficient fine-tuning
• Low-rank adapters with common prompt tokens