PromptGraph: Graph-Guided Prompt Sanitization for Balancing Privacy and Utility in LLM Inference

2026-07-12 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

PromptGraph is a novel graph-guided prompt-sanitization approach designed to balance privacy and utility during Large Language Model (LLM) inference. Addressing the challenge where sensitive information can be inferred from both explicit identifiers and contextual associations, PromptGraph explicitly models pairwise relationships among prompt spans, a limitation of existing sanitizers. It represents each prompt as an attributed graph, where nodes carry span-level privacy scores and edges encode contextual dependencies crucial for utility. The system's sanitization objective selects a protected span set to maximize privacy gain while minimizing the loss of these contextual dependencies. Protected spans undergo local sanitization, with placeholders restored only after passing local consistency checks. Experiments demonstrate that PromptGraph achieves a more favorable balance between privacy and utility compared to current prompt-privacy baselines.

Key takeaway

For AI Security Engineers or ML Engineers developing privacy-preserving LLM applications, PromptGraph offers a critical shift in how you approach prompt sanitization. You should move beyond simple span-level privacy signals and explicitly model contextual dependencies between prompt elements. This graph-guided method allows you to achieve a superior balance between protecting sensitive user data and maintaining the utility of LLM responses, preventing unintended information leakage from associations. Consider integrating graph-based analysis into your sanitization pipeline.

Key insights

PromptGraph balances LLM inference privacy and utility by modeling contextual dependencies between prompt spans via graph representation.

Principles

• Contextual associations contribute to privacy leakage beyond explicit identifiers.
• Explicitly modeling pairwise span relationships is key for privacy-utility balance.
• Graph representations effectively encode span privacy and contextual utility dependencies.

Method

Represent prompts as attributed graphs with privacy scores on nodes and contextual dependencies on edges. Select a protected span set to maximize privacy gain while penalizing contextual dependency loss. Sanitize locally and restore placeholders after consistency checks.

In practice

• Apply graph-based analysis to identify sensitive contextual dependencies in prompts.
• Design sanitization algorithms that optimize for both privacy and utility metrics.
• Incorporate local consistency checks for sanitized prompt restoration.

Topics

• Large Language Models
• Prompt Sanitization
• Data Privacy
• Graph Representation
• Contextual Dependencies
• Utility Preservation

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Disentangling Adversarial Prompts: A Semantic-Graph Defense for Robust LLM Security — A semantic-graph defense framework proactively disentangles adversarial prompt components to secure LLMs against jailbreaking and injection attacks.
• Long Context Isn’t Free — I Built a Safe Prompt-Pruning Layer That Makes LLM Systems Work — Deterministic, dependency-safe prompt pruning significantly reduces LLM context bloat without sacrificing critical information.
• PromptPrint: Behavioral Biometrics Through Natural Language Prompting in LLMs — Prompt-based identity, a new behavioral biometric, is primarily encoded in a user's surface-level lexical choices, not semantic intent.
• Formalizing and Mitigating Structural Distortion in LLM Attention for Zero-Shot Graph Reasoning — LLM performance on graphs is bottlenecked by structural distortion from sequence linearization, correctable by attention alignment.
• RAG Security and Privacy: Formalizing the Threat Model and Attack Surface — RAG systems introduce unique privacy and security risks requiring a formal threat model and specific defense strategies.
• From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model — LLM safety analysis benefits from paired prompt-to-response severity transitions and joint relevance assessment.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.