Prada enables private, efficient adaptation of large language models on resource-constrained devices without compromising security or performance.

• Privacy-First Approach: Keeps sensitive data local on user devices, eliminating privacy risks associated with data transmission
• Resource Efficiency: Operates within the limited computational capabilities of smartphones and personal computers
• Black-Box Method: Adapts LLMs without requiring access to model parameters or architecture details
• Practical Performance: Achieves comparable results to traditional fine-tuning while preserving privacy

This research addresses critical security concerns in AI personalization, allowing organizations to offer customized LLM experiences without exposing user data to third parties or requiring powerful hardware.

Prada: Black-Box LLM Adaptation with Private Data on Resource-Constrained Devices