I highly doubt this would work considering you train and sample over how common patterns and associations are. A unique pattern that is not dominating the datas

I highly doubt this would work considering you train and sample over how common patterns and associations are. A unique pattern that is not dominating the dataset will just be forgotten. Sampling especially will look to reduce these unique oddball responses. So unless we get a PoC with a trained model, the dataset used, the degree of poisoning and sampling explained, this is likely just fantasy. edit: if you downvote, please argue, for the benefit of us all. I work in the AI space, maybe there's something I missed.