Core Concepts
Innovative Sentinel-Guided Zero-Shot Learning paradigm facilitates collaboration without data exposure.
Abstract
The article introduces the Sentinel-Guided Zero-Shot Learning (SG-ZSL) paradigm to address data privacy concerns in AI collaborations. SG-ZSL utilizes a teacher model to guide student models without sharing sensitive data. Two training protocols, white-box and black-box, balance privacy and performance. Differential Privacy is integrated into teacher model training for enhanced security. Experimental results show SG-ZSL outperforms traditional ZSL methods in various tasks.
Stats
SG-ZSL consistently outperforms in ZSL and GZSL tasks.
White-box protocol enhances adaptability with distinct security-level training protocols.
Differential Privacy integrated into teacher model's training process.
Quotes
"Balancing the desire for openness with these issues is an ongoing struggle for researchers."
"SG-ZSL paradigm consistently outperforms in ZSL and GZSL tasks."