Protecting Personal Information using Homomorphic Encryption for Person Re-identification

We protect features extracted from person images with homomorphic encryption and compute distances between them without decryption. This allows people in different camera views to be matched while features containing personal information remain encrypted.

To check whether people in different camera views are the same individual, a computer compares features extracted from their images. These features can contain personal information, so they need protection when processed by an external server. We investigate how to compare them while keeping them encrypted.

We use homomorphic encryption, which allows calculations without first decrypting the data. Image features are represented as lists of numbers, and the computer calculates how similar two lists are while they remain encrypted. To make the features compatible with the encryption, we also convert their values into integers using a fixed rule. A server on a public network can then perform the comparison without decrypting the features.

We compared person-matching accuracy with and without encryption and the conversion to integers. Statistical tests, which take variation in the data into account, did not detect a clear difference in accuracy. This shows that the protected comparison could be used without a statistically detectable loss of matching accuracy in the experiments.

• Shogo Fukuda, Masashi Nishiyama, Yoshio Iwai,
  Reduction in Communication via Image Selection for Homomorphic Encryption-based Privacy-Protected Person Re-identification,
  Proceedings of 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), Vol. 5: VISAPP, pp. 36 - 47, February 2021. [Video]
• Kazunari Morita, Hiroki Yoshimura, Masashi Nishiyama, Yoshio Iwai,
  Protecting Personal Information using Homomorphic Encryption for Person Re-identification, Proceedings of IEEE 7th Global Conference on Consumer Electronics (GCCE), pp. 135 - 136, October 2018. [Student Paper Award]
• 福田 尚悟, 森田 一成, 西山 正志, 岩井 儀雄,
  準同型暗号を用いた人物対応付けのための画像選択による通信量の削減,
  電子情報通信学会論文誌 D, Vol.J103-D, No.10, pp. 721 - 732, October 2020.