TL;DR I posted a question yesterday; the answer to which gave me a new question:

Is Encrypted Intellectual Property still considered Intellectual Property in most (or, relevant) cases?

Here's some context for my particular situation:

I want to train a neural network to learn Dr. Seuss, whose works are copyrighted. That means, I would have to compile a large data set containing all or most of the works of Seuss, and store it all in a text file, or series of files. For simplicity's sake let's say 1 file containing all the books/poems concatenated together. All copyrighted, all intellectual property (IP) of someone else.

Here's a precursor question: Is it legal to even create that document, for personal/educational/fair use?

Assuming it is, even vaguely, then imagine a case where I create said document, and plug it into a neural network to generate Seuss-like works, and now I want to share the application on GitHub, publically, so other people can train their own Seuss neural network. That would mean sharing the IP dataset too, because I can't very well expect everyone who wants to clone my repository to generate their own Seuss dataset, because that's not a trivial thing to do.

The last question (which I linked above) told me that no, sharing it in this way is sketchy and ambiguous at best, if I wanted to claim fair use.

So then it became, "How can I do what I want to do, but without copyright infringement?" I had a couple of ideas on that, relating to encrypting the IP dataset.

One idea was to write a batch processing script that would do a simple en/decryption on a plaintext file, and include that with the repo.

Another idea was to encrypt the data initially before putting it in the repo, and include the decrypt script, as a subprocess to be run each time the neural network generates [samples] some original* data.

*original being used loosely here, as poorly managed neural nets are prone to overfitting, or providing samples which are not unique enough to be considered original data...

This was sort of a crazy idea to me, to think that I could potentially train a NN on encrypted (seemingly junk) data, and then when it generates its own gobbledygook, being able to run the decrypt script on it which would turn out to be accurate, legible, and coherent. But it lines up perfectly with what I am trying to do. (Is encrypted intellectual property still intellectual property in this case?)

Within the repo, the encrypted IP data is junk, meaningless and without context. Within the neural network, which would be training on the encrypted data, it's also arguably meaningless. At least, is not utilizing or infringing on any intellectual property in any straightforward way. Only when the user tells the neural network to generate some Seuss-like text, it spits out what it has learned. But what it has learned will be, for all intensive porpoises, an illegible layer of encrypted nonsense. Only after running the output through the decrypt will it even remotely resemble the works of Seuss.

So... what about all that? :) Does any of this resemble something legal? Or are they just pipe dreams with little basis in the reality of law?

Again, remember that due to overfitting, the sort of black-box system I've described could potentially output data that decrypts into verbatim Seuss intellectual property. But in reality, the level of 'closeness' to Seuss falls along a spectrum, which I informally call 'variable likeness.' If the sample is too exact, its overfit (which is a bad thing unless the user just wanted a really complicated program to print out exact Seuss stuff all along). On the other hand, if it is incomprehensible or nonsensical, it needs to keep learning, or start learning differently.

Also remember that the decrypt script would still exist somewhere in the repo, and someone could potentially use it or modify it to decrypt the actual IP dataset. But to be fair, that's their potential infringement, not mine.

Final question, which may seem weird: Does the copyright holder of Dr. Seuss' work also hold the copyright to anything resembling the 'variable likenesses' of Seuss as I have described? In other words, could decrypted samples generated from this network constitute Seuss IP, seeing as all the machine learned on was encrypted Seuss IP in the first place?

  • 1
    It would be pretty silly if giving someone an encrypted version of a file you aren't allowed to give them, and the decryption key, was allowed.
    – user253751
    Mar 10 '18 at 4:51
  • @immibis They'd never be decrypting the IP without some reverse engineering. They'd only be decrypting variable likenesses of the IP (samples the neural networks has produced after training on the encrypted IP data). I don't disagree with accuracy of your statement, only how closely it resembles the case in question.
    – Bango
    Mar 10 '18 at 7:07
  • Just entering the text into your system is copying. Copyright protects against copying. You are in the wrong before you even get to the distribution issues. Dec 1 '19 at 1:55

If it is fair use to copy it en clair then it is fair use to copy it enciphered. If it is not fair use en clair then it is not fair use enciphered.

Enciphered works are derivative works of the original and permission of the copyright owner or fair use is required to create the encipherment. Yes it is still IP.

  • Would it really be considered a derivative work if encryption is a one-to-one permutation? By that logic, printing something in a custom font would turn it into a derivative work.
    – forest
    Jan 14 '19 at 3:36

For your training data, standard intellectual property law applies. After all this is no different from an ebook.

A neural network trained on copyrighted data is probably a mess you don't want to get into. I doubt there is case law on it (didn't check).

The good thing is that the neural network probably can't reproduce significant verbatim snippets of the original text, so it's not a derivative work nor a verbatim copy and should be ok. Short phrases should not meet the requirements for copyright (say 3 or 4 words next to each other, even if the combination is fairly unique to the copyrighted work) and even some short quotes should fall under fair use.

But there is ongoing research on recovering training data from a neural network and if that becomes possible, then the neural network is just a verbatim copy with extra steps.

Also even if recovering the data is not possible, a sufficiently determined lawyer might come up with some other way to justify the claim that the neural network is a derived work or copy of the work.

  • Oops. I didn't see how old this question was before I answered. Well.
    – Nobody
    Nov 29 '19 at 21:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.