Face Anonymisation Hack Session

Algovera & DataUnion

Goal

• Create an algorithm to remove faces/personal information from images
• Apply this algorithm to all images uploaded to DataUnion’s image dataset
• The original images will be stored in the dataset but in the data portal (squid.dataunion.app) it will only show images with the filter/annotation on top of the images
• Put the algorithm in a continuous evolvement process via new uploaded data and via a challenge in the mobile app to fool the algorithm on purpose
• Distribute the value created by the algorithm to all the involved contributors (data scientists, data contributors, algorithm evolvers, facilitators)
• Offer the algorithm via compute-to-data as a service on Ocean Protocol, share the revenue made through sales, share the datatokens

How do we get there?

• Research which open datasets to use for a base version
• Balanced Faces in the Wild: https://github.com/visionjo/facerec-bias-bfw
• MIAP (More Inclusive Annotations for People) - https://arxiv.org/pdf/2105.02317.pdf: https://storage.googleapis.com/openimages/web/extended.html
• Research which algorithms are promising to use
• YOLOv5 - low resource consumption
• EfficientNet - https://github.com/google/automl/tree/master/efficientnetv2
• How will we allocate GPUs for the hackers?
• Discuss how we will share the results + the long term value of the results
• Create a shared GitHub repository + a GoogleDocs to start the work
• Start pragmatic and evolve from there!

Long term vision

• PlayAI mobile app - https://youtu.be/AsZUCkMZZm4?t=147
• Compute-to-data deployment
• Inclusion in the image dataset workflow (asynchronous calls of the algorithm after uploads)
• How do we keep the price realistic to use for a lot of images but still allow it to be a datatoken pool?