Article Preview
TopIntroduction
Who stole my “face”? Since the second half of 2018, the commercialization of face recognition has gradually been matured in all areas. According to the Freedom of Information Act of the United States, as early as 2014 the FBI collected photos of criminals’ faces to build a ‘new generation of identification database’ in order to quickly identify criminals. Major airports in Japan introduced face recognition systems in 2015 to provide faster inspection services for entry and exit personnel. FaceFirst's facial recognition system is deployed at airports and military bases in the United State. Domestically, face recognition technology has been fully applied in Alibaba’s Alipay mobile payment, face recognition classification security check of Beijing subway stations, residential quarters ‘door guards’ face recognition access control and other scenarios. The face recognition has become an indispensable part of people’s daily lives. However, in a risky society, the measures to prevent risks may bring new risks. In recent years, property and personal threats caused have emerged one after another because of data leakage, and the use of this technology by relevant departments may also bring new social risks.
“What it means when we become a commodity to be exploited. It’s similar to the right to exploit oil when David Carroll's personal information data being exploited as a commodity.” In May 2018, the European Union implemented the most stringent General Data Protection Regulation in history, stipulating the Internet companies which illegally collect personal information (including fingerprints, face recognition, etc.) and do not guarantee data security can be fined up to 2000 Million euros or 4% of global turnover. Nevertheless, on March 15, 2021, the China’s CCTV “3·15” exposed that more than 20 merchants across the country installed cameras to recognize faces and collected face data. Among them, Kohler (China) Investment Co., Ltd. installs face recognition cameras in its bathroom stores to capture personal information including gender and age. Its camera system can capture information without the customer’s knowledge and perception, and add various tags, such as professional fakers, reporters.
A typical face recognition system on the market first obtains a user's face image, and extracts high-dimensional face features through the pre-trained neural network. Then it stores the feature vector and identification tags in the cloud server. The front end of the system can recognize the identity, age, gender and even race of a person through these feature vectors. Therefore, the problem that how to maintain the performance of face recognition as much as possible while preventing the leakage of facial feature information has become the core issue of this paper.
An effective solution for the secure face recognition scheme is to use a cryptographic algorithm to encrypt the face template database and to match the cipher text data. However, this method doesn’t make some progress in the field of image recognition. Firstly, general encryption schemes do not support the basic operations required for feature matching in the encryption domain. Secondly, although the homomorphic encryption scheme supports some basic arithmetic operations on ciphertext data, it can only support a limited number of additions and multiplications, or the computational overhead is too large.
In 2017, Chen et al. (2017, December) proposed the CKKS homomorphic encryption scheme that supports real/complex approximate calculations. Because its ability to perform homomorphic encryption on floating-point numbers and its considerable computational efficiency, it become a hot spot in the cryptography field and is widely used in the near future.
Under the premise of protecting user privacy and security, face recognition is realized efficiently and securely in the cloud environment. This paper proposes a secure face recognition system based on the CKKS encryption scheme. In this system, the article uses the MTCNN algorithm (Zhang et al., 2016) to detect and correct face images, the FaceNet algorithm to extract the 512-dimensional feature vector of the face image, and the CKKS encryption scheme to encrypt the face features. Then the distance between the ciphertext feature vectors and the maximum value of the ciphertext distance is calculated. The secure face recognition finally realizes in the ciphertext domain, which avoids the infringement of the user's personal portrait rights and privacy.