Descriptor: Benchmarking Secure Neural Network Evaluation Methods for Protein Sequence Classification (iDASH24)

Area of Science:

• Computational biology
• Cryptography
• Machine learning

Background:

• Transformer models are increasingly used in bioinformatics, raising privacy concerns for sensitive genomic data.
• Secure evaluation of these models is crucial for privacy-preserving analysis.
• Existing benchmarks for secure model evaluation are limited.

Purpose of the Study:

• To introduce the iDASH24 homomorphic encryption track dataset for benchmarking secure evaluation of transformer models.
• To provide a standardized resource for testing privacy-preserving machine learning techniques.
• To facilitate the development of secure genomic data analysis methods.

Main Methods:

• Designed a dataset comprising a protein family classification transformer model and associated example data.
• Utilized homomorphic encryption schemes for secure model evaluation.
• Organized the iDASH24 Genomic Privacy Competition to test secure evaluation strategies.

Main Results:

• The iDASH24 dataset was successfully used in a competition setting for secure evaluation of a transformer model.
• Benchmarking results and companion methods were generated during the competition.
• The dataset facilitated the exploration of homomorphic encryption for secure machine learning.

Conclusions:

• The iDASH24 dataset is a valuable resource for benchmarking secure evaluation of neural network models, particularly transformers.
• It supports the advancement of privacy-preserving machine learning in genomics.
• Facilitates standardized testing of homomorphic encryption applications in AI.