> There was an attempt at poisoning training data with such invisible cloaking before and it turned out that this didn't really affect training quality Source?

> There was an attempt at poisoning training data with such invisible cloaking before and it turned out that this didn't really affect training quality Source? The authors claim this is not the case, and they have experiments showing it.