This research evaluates the efficiency of Differential Privacy techniques for protecting text data in LLM interactions, with a focus on utility preservation in resource-constrained settings.

• Demonstrates challenges in predicting utility of sanitized prompts, leading to wasted computing resources
• Analyzes the trade-offs between privacy protection and maintaining useful LLM responses
• Provides insights for organizations using pay-per-use LLM services to balance privacy and cost
• Highlights implications for security-focused implementations where computational resources are limited

For security professionals, this work offers practical guidance on implementing privacy protections that don't compromise response quality or waste resources when using third-party LLM services.

Preempting Text Sanitization Utility in Resource-Constrained Privacy-Preserving LLM Interactions