4 ms·
At the face of it, this seems like a fool's errand. The upper bound for obfuscation is that the obfuscated text should still be readable by a human with minima
by saimiam 10y ago
At the face of it, this seems like a fool's errand.
The upper bound for obfuscation is that the obfuscated text should still be readable by a human with minimal effort. To read, the average reader will looks for patterns like "replace j with the y sound." Once these patterns are determined, coding them into your NLP AI is trivial.
- emodendroket 10y agoIt doesn't have to be bulletproof to serve its purpose, as anybody who remembers downloading "Boon Joovi" MP3s from late Napster can attest. And if some transformations are ambiguous between multiple original spellings, so much the better.
- saimiam 10y agoSince the stated goal of the project is > The goal of the project is to make text hard to read for computers yet fairly easy to read for humans it has to be close to bulletproof. Humans should be able to decipher obfuscated text while computers should never be able to decipher the same obfuscated text. This is going to be impossible since AI is a fast follower to human ingenuity.
- emodendroket 10y agoThat's a pretty narrow reading of making it "hard for computers" to do something. Think of it this way: the security measures on most homes are completely inadequate to stop a determined attacker. Nevertheless, they work because most attacks are opportunistic.
- saimiam 10y agoThere are far, far, far more incompetent home burglars than competent ones. On the other hand, only a handful of AI's exist and by definition, they are getting better and better. So, using your analogy, when it comes to AI every attacker is a determined attacker so your security needs to be ~100% bulletproof.
- emodendroket 10y agoI think you're making a mistake in assuming sophisticated AI is going to be used everywhere. This will easily defeat simple word filters, for instance, which are still commonly used.
- kerbalspacepro 10y agoI think the idea is that if you have random obfuscation rules, a human can generalize almost immediately to read it. Like, you could string together paragraphs of different obfuscation rules and your brain would be able to switch between rules fairly quickly. The ML goal would be to be able to do the same thing.