Secure Deep Learning for Intelligent Terahertz Metamaterial Identification

Feifei Liu, Weihao Zhang, Yu Sun, Jianwei Liu, Jungang Miao, Feng He, Xiaojun Wu2020 · DOI

TL;DR

This paper introduces a novel approach for securely identifying metamaterials in mixtures using terahertz (THz) time-domain spectroscopy (TDS) combined with a crypto-oriented convolutional neural network (CNN) and homomorphic encryption (HE). This method achieves 100% accuracy while preserving data privacy, addressing a critical need for secure and intelligent material identification in the face of potential stealthy monitoring.

What Problem Does This Paper Address?

Metamaterials, engineered structures with unique electromagnetic properties, have a wide range of applications, from sensing and imaging to cloaking and telecommunications. As these materials become more integrated into various technologies, the ability to identify them, especially in complex mixtures, becomes increasingly important. However, distinguishing metamaterials from their constituent materials using traditional methods can be challenging due to overlapping spectral features and the influence of various experimental parameters like incidence angles and polarizations. Furthermore, the application of AI for material identification raises significant privacy concerns. If data is sent to a remote server for analysis, it could be exposed, leading to potential misuse or leakage of sensitive information. This paper tackles both the challenge of accurate metamaterial identification and the crucial need for privacy-preserving AI in this domain.

Methods

The core of this work lies in the integration of three key components:

  • Terahertz Time-Domain Spectroscopy (THz-TDS): This technique is used to capture the electromagnetic response of metamaterial samples. THz-TDS is chosen for its ability to extract both amplitude and phase information, which is crucial for material characterization.
  • Crypto-Oriented Convolutional Neural Network (CNN): A specialized CNN architecture is designed for metamaterial identification. This CNN is adapted for homomorphic encryption by avoiding operations that are not supported by HE, such as max-pooling. It uses convolutional layers followed by a "square activation" function for non-linear expression and a fully-connected layer for classification. The network is trained on raw frequency-domain data.
  • Homomorphic Encryption (HE): To ensure privacy, homomorphic encryption is employed. This allows computations (in this case, the CNN inference) to be performed directly on encrypted data. The data owner encrypts their THz signals using a public key and sends them to an untrusted server. The server runs the crypto-oriented CNN on the encrypted data, and the resulting encrypted output is sent back to the owner, who can then decrypt it using their private key. This ensures that the server never has access to the plaintext data or the identification results.

Datasets

The study utilized a dataset of 66 samples, comprising 32 samples containing metamaterials and 34 without. These samples were mixtures of metamaterials with background materials like glucose, lactose, and medicines. The metamaterial design parameters were consistent with previous research [2]. The dataset was augmented to increase its size and diversity. Data augmentation techniques included adding Gaussian noise with varying variances and attenuating the signals to simulate power reduction. The augmented dataset was then split into training and testing sets using 5-fold cross-validation.

Key Insights and Ideas

  • AI for Metamaterial Identification is Feasible and Superior: The research demonstrates that deep learning, specifically CNNs, can effectively learn discriminative features from raw THz spectra, outperforming traditional methods like SVM and human identification.
  • Privacy is Paramount for AI in Sensitive Domains: The integration of homomorphic encryption addresses a critical gap in applying AI to sensitive data, enabling secure identification services without compromising privacy.
  • Tailored CNN Architectures for HE: Designing CNNs that are compatible with HE is crucial. The paper highlights the need to constrain network depth and replace unsupported operations like max-pooling to ensure successful encrypted computation.
  • Data Augmentation Enhances Stability: Augmenting the THz data significantly improves the stability and convergence speed of the CNN model, making it more robust.

Key Results

  • 100% Accuracy: The proposed crypto-oriented CNN with HE backend achieved a remarkable 100% accuracy on the test sets, significantly outperforming human identification (56.95%) and traditional SVM (87.9%).
  • Efficient Inference: The CNN model, when run with HE, took 9.6 seconds to perform inference on 92 encrypted test signals. This efficiency is attributed to the chosen HE parameters (N=2¹³, L=8) and the parallel processing capabilities of the computation graph.
  • Privacy Preservation: The HE scheme ensures that the server performing the identification has no access to the plaintext data, guaranteeing privacy for the data owner.

Limitations and Future Work

  • Parameter Selection for HE: The paper notes that selecting appropriate parameters for HE is crucial to balance security and computational efficiency. While they found optimal parameters for their specific experiment, further research into HE parameter optimization for various applications could be beneficial.
  • Generalizability to Different Metamaterials: The proposed method was demonstrated on specific metamaterial mixtures. While the authors suggest the approach is applicable to other configurations with fine-tuning of the network structure, further validation on a wider range of metamaterials would strengthen its generalizability.
  • Computational Overhead of HE: Although the HE implementation is efficient, it still introduces some computational overhead compared to plaintext computation. Future work could focus on further optimizing HE schemes or exploring alternative privacy-preserving techniques to reduce this overhead.
  • Expanding THz Applications: The authors envision this work paving the way for broader AI applications in THz technology, including THz astronomy, security, and smart sensing, suggesting a rich area for future research.

Summary updated

BibTeX
@misc{liu2020secure,
  author = {Feifei Liu and Weihao Zhang and Yu Sun and Jianwei Liu and Jungang Miao and Feng He and Xiaojun Wu},
  title = {{Secure Deep Learning for Intelligent Terahertz Metamaterial Identification}},
  year = {2020},
  doi = {10.3390/s20195673},
}
Figure