DeepMarks: A Digital Fingerprinting Framework for Deep Neural Networks | Hacker News Reader

DeepMarks: A Digital Fingerprinting Framework for Deep Neural Networks

eprint.iacr.org

53 pointslainon10 comments

Is it just me, or does anyone else also think the very last the world needs right now is for the Deep Learning crowd to merge with the pro-DRM crowd? This seems even worse than copyright on databases and such...

EDIT maybe to clarify what I'm afraid of here: Copyright law is draconian these days, and as someone handling pre-trained models on a regular basis I am (was?) in the lucky position that all the models available are free, as is the software executing them. If this takes hold, those days are over and we might start seeing the kind of horror headlines that come with modern day arbitrary copyright enforcement...

Waaaaaaaay too late, I'm afraid... https://www.tineye.com/technology

Content fingerprinting has been around for ages, yes. This is model fingerprinting.

From the abstract:

> Sharing the trained DL models has become a trend that is ubiquitous in various fields.

Apart from the Model Zoo[1], are there any other sites for sharing serialized models?

[1] https://github.com/BVLC/caffe/wiki/Model-Zoo

TensorFlow recently announced TensorFlow Hub (https://www.tensorflow.org/hub/), which I think is similar.

it's not very organized for sure. but in research, barely anyone I know trains models from scratch just to use them afterwards. In pytorch and tensorflow, you'd not even have to go to a model zoo page to use the most common architectures, the library will just load it on demand into a folder in ~. So yes, using and sharing models is ubiquitous from what I can tell.

Yes, I used `load_data` and friends and it is certainly convenient. I was just curious about a kind of open source market place for models, where you could find niche models that people train in their spare time.

Because I see some people misunderstand the goal of this:

> DeepMarks introduces the first fingerprinting methodology that enables the model owner to embed unique fingerprints within the parameters (weights) of her model and later identify undesired usages of her distributed models

This is not a content fingerprinting technology, but rather a method for $google/$facebook to prevent/detect use of their models by others.

I wonder if you can leverage adversarial networks to eliminate the fingerprint while still maintaining model performance.

Looks like this has already been done: https://arxiv.org/abs/1802.04633

Reply on news.ycombinator.com