iDASH24 Secure Evaluation of Neural Networks for Protein Classification

To provide a standardized approach for testing and benchmarking secure evaluation of transformer-based models, we developed the iDASH24 Homomorphic Encryption track dataset. This dataset is centered on protein sequence classification as the benchmark task. It includes a neural network model with a transformer architecture and a sample dataset, both used to build and evaluate secure evaluation strategies. During the iDASH24 Genomic Privacy Competition, participants used this dataset to design secure evaluations of the neural network, where inputs were encrypted, and computations were carried out solely on encrypted data using Homomorphic Encryption schemes. Along with the benchmarking results and associated methods, the iDASH24 dataset serves as a valuable resource for evaluating secure neural network model evaluations.