kaylode / mediaeval21-drsf

De-identify facial information in image and video by using Adversarial Attacks

MediaEval 2021 - Driving Road Safety Forward: Video Data Privacy

Facial Data De-identification with Adversarial Generation and Perturbation Methods

Task Overview

• MEDIAEVAL21

Data Overview

• DATA

Method Overview

• Run 01: FaceSwap
  • Try swapping the face to fully hide the true identity of the drivers while preserving other information, such as facial expressions, behaviours, etc.
• Run 02: Adversarial Attack
  • While preventing and limiting unauthorized models from accessing and exploiting information of the de-identified data. The authorized model can access all information of the de-identified data.

Original image	Adversarial Pertubation	FaceSwap

Working note

• Our team's working note: HCMUS at MediaEval 2021: Facial Data De-identification with Adversarial Generation and Perturbation Methods