Scale-up Unlearnable Examples Learning with High-Performance Computing

Area of Science:

• Artificial Intelligence
• Data Security
• Machine Learning

Background:

• AI models like ChatGPT may inadvertently retain sensitive healthcare data.
• Medical imaging data used in AI diagnostics poses privacy and intellectual property risks.
• Unlearnable Examples (UEs) offer a novel approach to prevent deep learning models from learning specific data.

Purpose of the Study:

• To scale Unlearnable Clustering (UC) using high-performance computing (HPC) for improved UE performance.
• To investigate the impact of batch size on UE efficacy at HPC levels.
• To enhance data security and prevent unauthorized learning in AI models.

Main Methods:

• Utilized Distributed Data Parallel (DDP) training on the Summit supercomputer.
• Conducted experiments on diverse datasets (Pets, MedMNist, Flowers, Flowers102).
• Analyzed the relationship between batch size and unlearnability across different datasets.

Main Results:

• Scaling UC on HPC enabled exploration of UE performance with large batch sizes.
• Both overly large and small batch sizes negatively impacted UE performance and accuracy.
• The optimal batch size for unlearnability varied significantly across datasets.

Conclusions:

• Selecting appropriate batch sizes is critical for effective data protection using UEs.
• Dataset-specific batch size strategies are necessary for optimal unlearnability.
• HPC and DDP frameworks facilitate robust UE research for enhanced AI data security.