A Machine Learning-Based Secure Face Verification Scheme and Its Applications to Digital Surveillance

Abstract: Face verification is a well-known image analysis application and is widely used to recognize individuals in contemporary society. However, most real-world recognition systems ignore the importance of protecting the identity-sensitive facial images that are used for verification. To address this problem, we investigate how to implement a secure face verification system that protects the facial images from being imitated. In our work, we use the DeepID2 convolutional neural network to extract the features of a facial image and an EM algorithm to solve the facial verification problem. To maintain the privacy of facial images, we apply homomorphic encryption schemes to encrypt the facial data and compute the EM algorithm in the ciphertext domain. We develop three face verification systems for surveillance (or entrance) control of a local community based on three levels of privacy concerns. The associated timing performances are presented to demonstrate their feasibility for practical implementation.

• The paper introduces a secure face verification system using a DeepID2 CNN and homomorphic encryption to maintain privacy during identity verification.
• The proposed scheme achieves 96% accuracy on the LFW dataset while operating in near real-time with the Paillier cryptosystem.
• The adaptable system design supports varying privacy needs in surveillance, offering scalable encryption solutions and prospects for future feature extraction improvements.

Overview of a Machine Learning-Based Secure Face Verification Scheme

The paper presents a methodology for implementing a secure face verification system that uses machine learning techniques with substantial implications for digital surveillance applications. The authors focus on protecting identity-sensitive facial images from being compromised during verification processes. This research tackles the dual issues of maintaining verification accuracy while ensuring the privacy of facial data.

At the heart of this verification scheme is the DeepID2 convolutional neural network, which effectively extracts facial features from images. This is coupled with an Expectation-Maximization (EM) algorithm aimed at solving the face verification challenge in the encrypted domain. The authors integrate homomorphic encryption to facilitate computations on encrypted data, mitigating privacy concerns by preventing direct data access and thus safeguarding sensitive information.

Technical Contributions

1. Feature Extraction: The paper implements the DeepID2 CNN model to train facial image feature extraction. This extends beyond simple classification, allowing the model to derive meaningful facial features necessary for identity verification.
2. Secure Computation using Homomorphic Encryption: An essential contribution of the work lies in leveraging homomorphic encryption to encrypt facial data, permitting secure computations in the ciphertext domain. This encryption ensures that privacy is maintained even in the presence of potentially malicious system administrators or external threats.
3. Practical System Design: The authors describe the development of a face verification system suitable for surveillance or entrance control, catering to different privacy need levels through three proposed scenarios. These scenarios demonstrate varying degrees of data protection, from plaintext computation to full homomorphic encryption, thereby offering adaptable solutions depending on privacy requirements.

Results

The authors note notable accuracy from their proposed system, achieving a commendable face verification accuracy of 96% using the LFW dataset, which demonstrates the high efficacy of DeepID2 in feature extraction compared to traditional methods like HD-LBP. Furthermore, timing performance evaluations reveal that the system can operate in near real-time when configured with the Paillier cryptosystem for encryption, highlighting its potential for practical implementation in real-world scenarios.

Implications and Future Scope

This research provides a robust framework for secure face verification systems, crucial in domains such as digital surveillance where both accuracy and privacy are paramount. The adaptability of encryption methodologies—ranging from SFHE with the Paillier Cryptosystem to FHE—supports varied privacy settings, offering a scalable security solution.

The paper also indicates a forward-looking trajectory by implying future improvements in feature extraction precision through multi-patch DeepID network architectures and exploring methods to expedite encryption domain computations, especially within FHE frameworks. These pathways suggest continued refinement of both technical capabilities and practical applications in AI-driven security technologies.

Overall, this work provides a strong foundation for both theoretical exploration and practical implementation of secure, machine learning-based facial verification systems, contributing valuably to developments in the fields of cryptography, machine learning, and digital security.