FICGAN: Facial Identity Controllable GAN for De-identification

Jeong, Yonghyun; Choi, Joon-Seok; Kim, Sung-Won; Ro, Youngmin; Oh, Tae-Hyun; Kim, Doyeon; Ha, Heonseok; Yoon, Sungroh

doi:10.48550/arxiv.2110.00740

Cited by 3 publications

(5 citation statements)

References 40 publications

(62 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A key difficulty in these works has been generating high-quality images that capture useful information in the original image. To this end, recent works have focused on developing techniques to disentangle and preserve non-identity attributes of the image, such as pose and facial expression [21,22,34,58]. However, these methods are not directly applicable to our setting given the unclear distinction between identity vs. non-identity features in retinal images beyond the blood vessel structure.…”

Section: Related Workmentioning

confidence: 99%

k-SALSA: k-anonymous synthetic averaging of retinal images via local style alignment

Jeon,

Park,

Kim

et al. 2023

Preprint

View full text Add to dashboard Cite

Section: Related Workmentioning

confidence: 99%

k-SALSA: k-anonymous synthetic averaging of retinal images via local style alignment

Jeon,

Park,

Kim

et al. 2023

Preprint

View full text Add to dashboard Cite

“…This section compares and evaluates the privacy and utility of the proposed CFIDM with APFD [4], k-Same-net [5], KSS-GAN [6], KSS-APFD [7] and FICGAN [8] through experimental simulation evaluation.…”

Section: Experimental Analysismentioning

confidence: 99%

“…Yan et al [7] used the ELEGANT model to encode original face attributes and synthesize them into K-selected face images to ensure the anonymity of face images to be recognized. To realize face multi-attribute derecognition, Jeong et al [8] decoupled images with encoders and protected identity-related attributes with the k-SAME mechanism. Controllable face identity de-recognition was achieved by adjusting the tradeoff between de-recognition degree and reserved facial attributes.…”

Section: Introductionmentioning

confidence: 99%

Controllable identity de-identification for face image

Shen

Zhang

2022

International Conference on Computer, Artificial Intelligence, and Control Engineering (CAICE 2022)

View full text Add to dashboard Cite

Aiming at the problem that the face image recognition technology based on anonymization mechanism cannot provide quantifiable privacy guarantee and realize controllable privacy protection, a controllable face identity recognition method based on differential privacy mechanism and generative learning model was proposed. Based on the attentional mechanism and graph convolution network, this method designed a representation learning model to extract identity-related geometric and demographic information by exploring spatial and spatial semantic relationships between different facial regions. The differential privacy mechanism was used to protect the demographic attributes while preserving the geometric attributes of the face. Then generate adversarial network reconstruction image instead of original image release for computer vision analysis task. Experimental results show that the proposed method is superior to other models in terms of privacy and utility. It can effectively resist background knowledge attacks, reconstruction attacks, and combination attacks while realizing verifiable and controllable privacy protection of face identity and ensuring the optimal tradeoff between privacy and utility of published images.

show abstract

“…pose, glass, hairstyle, elevation, etc.). Recent works [2,5,26,45] focus more on identity and non-identity disentanglement. Specifically, Nitzen et al [45] propose a latent space mapping network to map both the identity and attribute representations into StyleGAN latent code.…”

Section: Related Workmentioning

confidence: 99%

“…Likewise, because of the limited information contained in StyleGAN latent code, this approach cannot preserve all the non-identity attributes, especially in hair and background. To better preserve the nonidentity attributes, FICGAN [26] uses a much larger latent code to better restore expression and pose. However, the facial identity is not fully disentangled, meanwhile, the hair and background details are not totally preserved either.…”

Section: Related Workmentioning

confidence: 99%