Anyone have tips on detecting whether a sequence of bytes was generated by a tool like GPT-3? Not necessarily just for plagiarism detection, I'm more interested in things like detecting malware and trojans.
I’ve been thinking about this lately, and I am just getting into AI and NNs so take this with some salt…
A service could easily add something after the fact - like a water mark, but I think if it was mixed in with the content it would get learned and washed out when training or validating the output.
For something more fun you might be able to build a classifier per student that can determine if a student made the text. You’d have to build it based on the students past work - you could then get the probability that that student wrote that paper.
That classifier could also, maybe, be able to find the “voice” of gpt3 vs something else.
You could game that system by using AI from the start, or slowly adding in more and more AI paragraphs, in say, high school - where most of the training data for a student would come from.
On the flip side a clever student could build their own “AI based content generator” (not sure what to call that) and use their own content - which is my latest little side project.
I'm interested in identifying the tool used to generate the text.
For example, recently I used a word2vec-like approach on obfuscated assembly code. With a relatively small model this can get reasonable answers to questions like "which of these obfuscations was done to the code" and "which of these un-obfuscated functions is the best match for this obfuscated code".
So I am wondering what approaches people are using to detect when media was generated by a machine.
My research interest is in the binary/malware niche, but I'm using work from NLP and computer vision. And I figure other people are interested in answering "was this artwork made by a human" or things like "was my copyrighted artwork used to generate this AI-generated image".
