Separable Expert Architecture: Toward Privacy-Preserving Large Language Model Personalization via Composable Adapters and Deletable User Proxies

The paper proposes a three-layer architecture that decouples users' personal data from shared weights, consisting of a static base model, composable domain-expert LoRA adapters that carry no user data, and per-user proxy artefacts whose deletion constitutes deterministic machine unlearning. Evaluations on Phi-3.5-mini and Llama-3.1-8B confirmed per-user differentiation, with personal data remaining isolated while influencing outputs; after deleting the proxy, outputs returned to baseline (KL divergence of about 0.21 nats, 82-89% validation pass rate), and cross-user contamination was near zero. Because user-specific information never enters the shared weights, the architecture mitigates model inversion, membership inference and training data extraction against shared model components by construction, and turns machine unlearning from an intractable weight-editing problem into a deterministic deletion operation, compatible with DP-SGD for privacy-preserving improvement of the shared model.