SecureLLM: Using Inference-time Compositionality to Build Secure Language Models

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

SecureLLM is a novel compositional framework designed to enhance data confidentiality and access control in Large Language Models, particularly for critical sectors such as healthcare, finance, and public governance. Addressing the limitations of existing LLM safety approaches in strictly segregating confidential data, SecureLLM integrates fine-tuning with traditional access security measures. The framework operates by fine-tuning LLMs on segregated, "siloed" training data and then composing their outputs at inference time, strictly based on a user's verified credentials. This methodology prevents unauthorized data leakage while still allowing accurate responses to complex queries that span multiple data silos. Its effectiveness is demonstrated on a challenging natural-language-to-SQL translation task, highlighting its applicability in real-world scenarios where sensitive information protection is paramount.

Key takeaway

For AI Security Engineers designing LLM deployments in sensitive sectors like healthcare or finance, SecureLLM offers a robust approach to enforce data confidentiality. You should consider implementing inference-time composition with fine-tuned, siloed models to prevent unauthorized data leakage while still supporting complex, cross-domain queries. This method allows you to maintain strict access control based on user credentials, mitigating significant privacy risks.

Key insights

SecureLLM secures LLMs by fine-tuning on segregated data and composing outputs at inference time based on user credentials.

Principles

• Data segregation improves confidentiality.
• Inference-time composition enforces access.
• User credentials gate data access.

Method

SecureLLM fine-tunes LLMs on segregated data silos. At inference, it composes outputs based on verified user credentials, preventing unauthorized data leakage while enabling complex cross-silo queries.

In practice

• Secure natural-language-to-SQL translation.
• Protect healthcare data confidentiality.
• Safeguard financial sector information.

Topics

• SecureLLM
• Large Language Models
• Data Confidentiality
• Access Control
• Inference-time Composition
• Natural Language to SQL

Best for: AI Architect, Research Scientist, CTO, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• SPARK: Security Knowledge Priming and Representation-Guided Knowledge Activation for LLM-based Secure Code Generation — SPARK activates latent security knowledge in LLMs at inference time using prompt cues and logit biasing, avoiding retraining.
• Security Risks in Language Models (LLMs) — LLMs face unique security threats from manipulation, data leakage, and supply chain vulnerabilities across their lifecycle.
• CDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipes — Current LLMs struggle with the procedural faithfulness required for compositional, order-sensitive data refinement recipes.
• The ACUTE Protocol: Operationalizing Language Model Activations for Better Calibration, Utility, and Trust — The ACUTE Protocol enhances LLM trustworthiness by balancing calibration and informativeness through activation-based confidence estimation.
• What If LLM Workflow Could Be Orchestrated Like Writing SQL? — SPL orchestrates LLM workflows like SQL, separating "what" from "how" and integrating deterministic verification.
• Beyond Refusal: Probing the Limits of Agentic Self-Correction for Semantic Sensitive Information — Agentic rewriting effectively mitigates semantic sensitive information in LLMs while preserving utility, outperforming refusal-based defenses.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.