90 points by ml_magician 5 months ago flag hide 18 comments
gnosis 5 months ago next
Fascinating! I've been following biometrics closely lately, and your ML-based approach seems to address a number of shortcomings in traditional systems. Looking forward to exploring this!
ruby_beginner 5 months ago prev next
Nice! Do you think this approach would also work for fingerprint authentication on mobile devices?
gnosis 5 months ago next
Interesting thought. Maybe an optimized version due to the low dimensional data could perform well. However, you might still face some security issues since it's a different data distribution when compared to the facial dataset.
the_jethro 5 months ago prev next
Indeed, I'm worried about the privacy concerns surrounding facial recognition data. How do you address those?
gnosis 5 months ago next
Excellent question. Our system limits the amount of data exposure by using differential privacy and secure enclaves for computations. Users are also provided with options for encrypted data transmission as an added level of protection for their information.
tobold 5 months ago prev next
What ML techniques have you used for this? Any technical paper or repository I can look into for details?
gnosis 5 months ago next
We relied mostly on deep neural networks with Batch Normalization and Xavier initialization. Our training process included both supervised and semi-supervised learning methods. Detailed explanations and the code are available on our GitHub repository: <https://github.com/gnosis/ml-biometric-auth>
uga_jimmy 5 months ago next
I took a brief look at the code. You're using ResNet50 as a backbone architecture. Have you tried incorporating SqueezeNet or MobileNet to reduce the number of parameters or Inception networks to capture information from different scales?
gnosis 5 months ago next
Thanks @uga_jimmy. We've tried using SqueezeNet and MobileNet but faced some issues with the lower representational power. We did try and incorporate Inception networks and found that it improved our results. Will share those results soon in the repo.
swindol 5 months ago prev next
Have you thought about incorporating GANs to augment the dataset and make the classifier more robust to different lighting conditions and facial attributes? That could be a really interesting approach!
gnosis 5 months ago next
Thank you for the input. Yes, we initially experimented with GANs; the results were promising. However, we realized that the computational requirements along with stability concerns prevented us from generalizing it to different scenarios. Inspiring idea, nonetheless.
kadie 5 months ago prev next
Did you have to build the system with real-time performance in mind? I suppose it depends on the size of the DL models being used. Could deep regression networks be fine in this case?
gnosis 5 months ago next
Great question. The real-time performance was one of our main requirements. Deep regression networks, given their smaller size, could be considered, but they usually sacrifice a little accuracy compared to the larger models.
sw519 5 months ago prev next
@gnosis has open-sourced this project, so I guess we can all contribute to making it more accurate and stable, which is fantastic.
goofy_king 5 months ago prev next
Have you also considered OS-based multi-factor authentication as a more secure authentication mechanism?
gnosis 5 months ago next
Indeed, combining different factors of authentication does significantly increase the system's overall security. Our main aim was to create a user-friendly authentication method, and biometrics proved ideal with that in mind; however, a hybrid approach would be optimal for more secure scenarios.
interstellar_traveler 5 months ago prev next
Fascinating work, been working on a similar system for forensics and law enforcement purposes. Are there any ethical concerns, such as masked individuals opening somebody else's account or false positives?
gnosis 5 months ago next
Great question. Yes, we've faced concerns about masked individuals and false positives. We're currently allocating resources to improve detection rates when it comes to masked cases while also implementing multiple checks to mitigate false positives.