> And in fact I've never stated that the "model" contained the whole source dataset. Apologies for seeming rude, but I feel that the abstract is disingenuous (

> And in fact I've never stated that the "model" contained the whole source dataset. Apologies for seeming rude, but I feel that the abstract is disingenuous (I'm assuming you're the author of the article). The abstract states: > we provide a ... decompiler that reconstruct the source dataset from the cryptic matrices that constitute the software executed by them. But that's not what's happening here. Instead, what's happening is (correct me if I'm wrong) the decompiler uses the gradient information along with the network itself (which is very close to the penultimate network) to reconstruct the input. If we consider for instance, MSE loss, that reconstruction appears trivial given all the information available. As I said before, this reconstruction (while interesting) does not show copyright violation, because the training process information is not available once the network is deployed. Obviously the model contains information about its training set. If it didn't, it would be useless. I'm not saying there aren't obvious copyright issues, and I'm also not saying that the approach to recover the training set is not interesting. I'm just saying that the overall copyright argument has a major gap (and there are more direct alternative arguments).