Face Image De-identification Using Emoji Conversion for Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition technologies using deep learning face a privacy invasion issue when face images are exposed, as existing masking or encryption techniques require frequent decoding and are not fundamentally secure, making them impractical for large-scale learning databases.
Innovation Solution
A face image de-identification apparatus and method that generates an emoji image retaining facial feature information using an emoji conversion deep learning model, specifically a generative adversarial network (GAN), which replaces the original face image in the database, ensuring the identity cannot be visually distinguished while maintaining feature information for learning purposes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If masking or encryption techniques are applied to face images, then privacy protection is improved, but the face image cannot be fundamentally secured because unmasking/decoding is required for every learning event, making real-time processing slow and impractical
Solution Approach 1:
The patent applies preliminary action by converting face images to emoji images during the initial database construction phase. The emoji conversion deep learning model pre-processes face images into emoji representations that retain facial feature information but obscure identity. This one-time conversion during database creation eliminates the need for repeated decoding during subsequent learning events, thereby improving both privacy protection reliability and real-time processing productivity.
2Reliability
If masking or encryption techniques are applied to face images, then privacy protection is improved, but the processing time increases significantly when unmasking/decoding is required for every learning event
Solution Approach 1:
The patent performs the face image conversion to emoji image during the initial database construction phase rather than during each learning event. The emoji conversion deep learning model pre-converts face images into emoji representations that preserve facial features while protecting identity. This preliminary conversion during database creation eliminates repeated decoding operations during learning, significantly reducing processing time loss while maintaining privacy protection.
3Adaptability or versatility
If face images are used as learning database, then deep learning face recognition technology can be developed, but privacy invasion problem occurs when the database is exposed
Solution Approach 1:
The patent applies copying by creating emoji images that replicate the essential facial feature information of original face images while eliminating identifiable personal characteristics. The emoji conversion deep learning model generates emoji representations that contain the structural and geometric features needed for face recognition training, but do not reveal the actual identity of individuals. This copying approach allows the database to be exposed or shared without causing privacy invasion, while still enabling deep learning face recognition technology development.
4Reliability
If emoji images are generated using deep learning model, then facial feature information is retained while identity is obscured, but the model complexity increases
Solution Approach 1:
The patent introduces emoji images as an intermediary representation between original face images and learning database. The emoji conversion deep learning model acts as a mediator that transforms face images into emoji images, preserving facial feature information while obscuring identity. This intermediary emoji representation layer allows the system to maintain privacy protection through the emoji-ification process, while the model complexity is managed through the use of pre-trained emoji conversion models that can be applied efficiently during database construction.
Data Source
AI summary
A face image de-identification apparatus and method are disclosed. The face image de-identification apparatus may include an emoji generator configured to generate a first emoji image including facial feature information corresponding to a face image using the face image stored in a database, and an image inserter configured to insert the first emoji image into the database by replacing the first emoji image with the face image.