Secure Deep Learning for Intelligent Terahertz Metamaterial Identification

Abstract

Metamaterials, artificially engineered structures with extraordinary physical properties, offer multifaceted capabilities in interdisciplinary fields. To address the looming threat of stealthy monitoring, the detection and identification of metamaterials is the next research frontier but have not yet been explored. Here, we show that the crypto-oriented convolutional neural network (CNN) makes possible the secure intelligent detection of metamaterials in mixtures. Terahertz signals were encrypted by homomorphic encryption and the ciphertext was submitted to the CNN directly for results, which can only be decrypted by the data owner. The experimentally measured terahertz signals were augmented and further divided into training sets and test sets using 5-fold cross-validation. Experimental results illustrated that the model achieved an accuracy of 100% on the test sets, which highly outperformed humans and the traditional machine learning. The CNN took 9.6 s to inference on 92 encrypted test signals with homomorphic encryption backend. The proposed method with accuracy and security provides private preserving paradigm for artificial intelligence-based material identification.

Keywords

• metamaterial identification
• deep learning
• homomorphic encryption
• private preserving
• terahertz time domain spectroscopy (THz-TDS)