Date: Wed, 25 Aug 93 08:37:07 -0700 From: strick -- henry strickland Extracted from "FC NEWSBYTES 1.3", David Geddes Editor, where FC = FutureCulture mailing list . strick ____________________________________________ _ _.......... B Y T E 4: Visa, HNC Inc. develop neural network as a weapon to fight fraud SAN FRANCISCO (AUG. 10) PR NEWSWIRE - Visa International and HNC Inc. have announced a strategic agreement to develop a comprehensive merchant risk detection system. The new system will be designed to better control fraud at the merchant level by determining the risk associated with individual card transactions. For those who are not familiar with the details of neural networks, I thought I would point out that this represents a departure from the current notion of a credit rating in two ways: 1) There is no clear way to fix your "neural credit rating" if there is a problem. The neural network program which predicts the probability of fraud will give its guess as to the probability of fraud. If you are a cardholder and it predicts that a transaction is likely to be fraudulent, then your purchase won't be accepted. But, unlike conventional credit reporting firms which use a credit report, the neural network cannot explain anything about how it came to its decision. With existing credit reporting schemes, you at least have the option of acquiring your credit report and taking the necessary steps to repair your credit rating if there is a problem. With the use of neural networks, this is no longer possible. Given the current state of neural network research, a percentage of the rejections will be false. This means that a number of card users will be denied service for no other reason than the fact that neural networks make mistakes. 2) You are no longer judged on your own actions, but on the similarity of your purchasing patterns with those who have committed fraudulent acts. Instead of being judged on your trustworthiness based on your past actions, you will be judged based on whether people whose purchasing profiles are similar to yours are trustworthy. An example of this being problematic is say you purchase a particular CD and the neural network decides that, partly based on this and partly on other information, that you won't pay your bill because most of the people in the database who bought that CD didn't pay their bills. Andy

From: bill@twwells.com (T. William Wells) Date: Wed, 25 Aug 1993 21:04:05 GMT In article <9308252001.AA14017@custard.think.com>, Andy Wilson wrote: : [mostly bogus stuff] That is irrelevant to cypherpunks, as I understand the list. The prospect of the impossibility of anonymity and the uses to which personal information is made in a cashless economy is not relevant? I beg to differ. This is exactly what digital cash is meant to prevent. There is no technology, including that of privacy, that cannot be used for ill. We don't know how they're going to be using the neural network. They could, as was suggested, abandon their minds and and rely on the neural net. I don't think they will because doing so would be a really bad business decision. Furthermore, on the evidence, the neural network output will only be used as one datum in a process involving many inputs and a human making the final decision. Finally, in the examples I'm familiar with (from reading AI Expert), when a neural net is used as a decision element, precisely because of its error rate, the decision isn't "go/no go" but "go/refer the problem to a human". The problem with referring a neural network's decision to a human is that the neural network gives no information other than the probability of fraud. It does not tell the human why it determined the transaction was likely to be flawed, like a system based on rules or case-based reasoning would be able to do. There is not any good way to combine the judgement of the neural net with that of a human for that reason. With respect, I have found AI Expert to consist more of marketing hype than correct and useful information on artificial intelligence technology. Andy

From: bill@twwells.com (T. William Wells) Date: Thu, 26 Aug 1993 01:21:03 GMT [...] But that wasn't what you were writing about. You were writing about bad business decisions, not violations of privacy. No, you were writing about bad business decisions. I was providing a few details on how credit/charge-card information is used in this process and a few potential problems resulting from it. For that matter, your notions on neural networks seem contradictory. On the one hand, you complain about a violation of privacy and on the other you complain that a neural network won't tell you how it reached its conclusions! You are deliberately confusing two different points: 1) the fact that neural networks do not provide useful explanations of how they arrived at a particular decision, and 2) some potential problems that arise from this fact that concern privacy issues. : I beg to differ. This is exactly what digital : cash is meant to prevent. Digital cash and the use of neural networks to authenticate transactions are essentially orthogonal issues. I will reiterate that the whole point of digital cash is to provide anonymity, which will prevent these kinds of uses made of personal information which are not done with the explicit approval of the person involved. : The problem with referring a neural network's decision to a human : is that the neural network gives no information other than the : probability of fraud. 1) This statement is false. It is true of some neural networks but not all. We have no way of knowing whether their neural network is among those. It is true of all commercial applications of neural networks to my knowledge, and certainly true of the neural networks developed by Hecht-Nielsen. : There is not any : good way to combine the judgement of the neural net with that of a : human for that reason. Nonsense. As the existence of rule based systems that incorporate neural networks shows. That shows no such thing. The only way to combine the judgement of a neural network with that of a rule-based system, or anything else, is to see if both arrive at the same conclusion. You cannot see the reasoning process of the neural network to help the human understand why it made the judgement that it did, the marketing hype of neural network vendors notwithstanding. This is my last post on this thread. Andy