(As reported on KHON-TV2). The Hawaii Republican Party was up in arms today when it was discovered that a "hard disk crash" caused a number of absentee ballots on the island of Maui (state of Hawaii) to inadvertently omit two candidates running for state legislature. The crash, which reportedly occurred at a California company which had been contracted to print the ballots, means that over 140 Maui residents are now being asked to go in and re-cast their ballots. This printing error was discovered when friends couldn't find their friend's name on the ballot. The security risks are obvious--if the printing of ballots is so dependent on a single hard drive, it would seem fairly straightforward that people could sabotage elections this way. Lani Teshima-Miller (firstname.lastname@example.org.Hawaii.edu) "Sea Hare" UH School of Library & Info Studies.
An AP item from 15 Sep 1994 (e.g., San Francisco Chronicle, D3) noted that the Federal Trade Commission filed a false-advertising complaint against Brian Corzine of Sacramento, California, for making false claims in promoting a credit-repair program. A federal court ordered the promotion stopped. Corzine maintained he was merely a reseller.
In his famous paper "The relation of habitual thought and behavior to language" (published in his collected papers, "Language, Thought, and Reality" by MIT Press), the American linguist Benjamin Lee Whorf began with some tales from his day job as an insurance risk expert, in which he described the lapses of reasoning that led to accidents that his company had to pay for. He says, for example, that people often assume that an "empty barrel" is wholly innocuous -- even if it recently contained gasoline and can thus be reasonably expected to be exhaling flammable vapors. He suggests that part of the problem is with the word "empty", which derives from a particular folk model of objects and materials that unfortunately does not correspond accurately enough to reality. I am sure that Whorf is cringing in his grave at the power outage that hit O'Hare Airport this past Wednesday morning, which is described in a brief article in the Wall Street Journal: Daniel Pearl, A power outage snarls air traffic in Chicago region, Wall Street Journal, 15 September 1994, page A4. It would seem that power was lost for two hours when someone shorted out the power system for the Aurora air traffic control center while testing the uninterruptable power supply. "There are certainly some additional precautionary measures that have to be taken to make sure this doesn't happen again", said Stan Rivers, the FAA's deputy associate administrator for airway facilities. One precaution, he said, is not to perform the work when air traffic is at its peak. The power outage occurred at 8:45am CDT at the Aurora, Ill., facility, which serves traffic to and from the world's busiest airport, Chicago O'Hare International Airport. Now, I have no first-hand knowledge of this event, but I do think it is legitimate to ask what could possibly have possessed someone to mess with the power supply for O'Hare Airport's traffic control at 8:45 in the morning. I would like to suggest that, just maybe, it has to do with the phrase "uninterruptable power system". Says the Journal, It was the second time this year that the installation of an "uninterruptable power system" interrupted power at an air-traffic center. Granted the first occurrence was more freakish than the second (a falling ladder a hit "break glass and hit button" button). But do you suppose that these things would happen during the morning rush hour, as opposed to 4AM, if uninterruptable power supplies were called "back-up kludges for highly sensitive and fragile power supplies"? Where do hubristic terms like "uninterruptable" come from? They come from a very narrow understanding of "the system" -- if the power supply can overcome those mishaps that can happen within the narrow technical conception of the system that is found in box-and-arrow type diagrams, then it's "uninterruptable" so long as the world outside remains totally pure and safe. And if you don't believe me, I've got an "inherently safe nuclear reactor" to sell you -- I kid you not! Phil Agre, UCSD
Dennis Lawrence reported on an ad from TigerDirect, Florida, offering a set of 650 high-quality logos of major corporations. Logos are copyright by their owners and in many cases trade marked. This means that the copyright owner has sole rights to control who gets copies and who distributes. If TigerDirect has the explicit permission of the owners of the logos, all is well. If not, then not only they, but anyone else using the logo without authorization, is breaking the law. Anyone who would use a logo, authorization or no, to commit a fraud is also breaking the law. Peter Denning
100 Reasons to Oppose the FBI Wiretap Bill Reason 55: The largest purchaser of telecommunications equipment in the federal government said the FBI wiretap plan would have an *adverse impact* on national security. In 1992 the General Services Administration wrote that the FBI wiretap plan would make it "easier for criminals, terrorists, foreign intelligence (spies) and computer hackers to electronically penetrate the phone network and pry into areas previously not open to snooping." The confidential memo was obtained as a result of a Freedom of Information Act request. What To Do: Fax Rep. Jack Brooks (202-225-1584). Express your concerns about the FBI Wiretap proposal. 100 Reasons is a project of the Electronic Privacy Information Center (EPIC) in Washington, DC. For more information: 100.Reasons@epic.org.
Reading Peter Ladkin's account of his colleague's address woes brought to mind an incident that occurred to my friend recently. He moved and sent an address correction to a company in which he holds some stock. The company acknowledged his change of address, but sent it to his *old* address. One wonders if the database update occurred after the letter was generated or if it never happened. He has not received any mail since. Paul T. Keener email@example.com
The following is the abstract of a paper that will be presented at ESORICS 94. The full text is available via anonymous ftp from ftp.uni-hildesheim.de in /pub/publications/Sirene/publications, file BBCM1_94CafeEsorics.ps. The ESPRIT Project CAFE: High Security Digital Payment Systems Jean-Paul Boly, Antoon Bosselaers, Ronald Cramer, Rolf Michelsen, Stig Mjolsnes, Frank Muller, Torben Pedersen, Birgit Pfitzmann, Peter de Rooij, Berry Schoenmakers, Matthias Schunter, Luc Vallee, Michael Waidner CAFE (Conditional Access for Europe) is an ongoing project in the European Community's ESPRIT program. The goal of CAFE is to develop innovative systems for conditional access, and in particular, digital payment systems. An important aspect of CAFE is high security of all parties concerned, with the least possible requirements that they are forced to trust other parties (so-called multi-party security). This should give legal certainty to everybody at all times. Moreover, both the electronic money issuer and the individual users are less dependent on the tamper-resistance of devices than in usual digital payment systems. Since CAFE aims at the market of small everyday payments that is currently dominated by cash, payments are offline, and privacy is an important issue. The basic devices used in CAFE are so-called electronic wallets, whose outlook is quite similar to pocket calculators or PDAs (Personal Digital Assistant). Particular advantages of the electronic wallets are that PINs can be entered directly, so that fake-terminal attacks are prevented. Other features are: * Loss tolerance: If a user loses an electronic wallet, or the wallet breaks or is stolen, the user can be given the money back, although it is a prepaid payment system. * Different currencies. * Open architecture and system. The aim is to demonstrate a set of the systems developed in one or more field trials at the end of the project. Note that these will be real hardware systems, suitable for mass production. This paper concentrates on the basic techniques used in the CAFE protocols. Michael Waidner, Universit"at Karlsruhe (currently on leave to IBM Zurich Research, email: firstname.lastname@example.org)
BKFUZHBK.RVW 940616 Academic Press, Inc. 955 Massachusetts Avenue Cambridge, MA 02139 Josh Mills, Marketing, email@example.com firstname.lastname@example.org "The Fuzzy Systems Handbook", Cox, 1994, 0-12-194270-8 We dinosaurs of procedural programming language orientation tend to have problems with functional programming languages such as Prolog. It is difficult to reorganize your thinking into the existential model of the expert systems programmer. We have similar problems with object oriented programming. The difficulties we have with fuzzy logic are probably for the same reason. Take heart, fellow dinosaurs. At the very least, this book explains *why* we find fuzzy systems so troublesome. They are simply expert systems with a better conceptual grasp of probabilities. The trade media has hyped fuzzy systems as the new and coming thing. Information systems professionals, who have lived through a great number of "coming things" still know little more, basically, than the fact that control systems are supposed to be better with fuzzy logic, and that close now counts in both horseshoes and fuzzy systems. The reason for the confidence in this science of imprecision can, in part, be demonstrated in the control realm. Suppose you are building an automatic collision avoidance system for cars. It is fairly straightforward to program in the sequence of actions to be taken to slow the car as it approaches another object. If, however, the system fails, then what happens if you do hit the object? Will the "distance" become a negative number? If so, will the brakes bind or release? Will the drive train stop, accelerate, or go into reverse? This situation is simplistic, but the outcome, in a procedural language, must be accounted for, prepared and tested. Fuzzy systems deal with ranges, and it is much easier to see and understand that the concept of "close" should also include "hit" -- even before you start to build the actions to be taken. The potential disasters associated with systems that would flip planes upside down when they flew over the equator are not confined to control systems, as IS professionals are all too well aware. Financial disasters can be precipitated by "decision support" software which can generate market crashes. Similar damage can be done on a smaller scale by specialized programs which may have undiscovered, and unintended, assumptions. As with control systems, working with ranges may make the pitfalls more obvious than working with static and sterile values. Even so, it is difficult for the programmer to translate the concepts of fuzzy logic into code to play with. Cox has, therefore, given you code to play with. A high density MS-DOS format floppy contains C++ source code to mess around with. (C programmers should be able to work with most of it, and, since it is source code, Mac devotees should be able to use it, as well.) For those wishing to explore this new field "hands-on", a slightly high-toned, but very useful, introduction. copyright Robert M. Slade, 1994 BKFUZHBK.RVW 940616
I have attended several presentations from neural net software vendors. If these sales pitches are any indications, then neural nets are being used for just the opposite of the dark purposes Fred Baube fears. Successful lenders (and insurance companies) are driven by two basic business rules: avoid "bad" risks and sell more product (increase market size). Very broadly applied underwriting rules (don't write mortgages in ZIP codes with high default rates, don't insure Corvettes driven by young males) support the first rule but run counter to the second. There is only so much market share that can be taken from the competition. What used to be considered a marginal market needs to be reconsidered in order to fuel the business growth that Wall Street demands. Many companies are essentially saying: "on first blush you sound like a bad risk - find me an excuse to sell to you" - and neural nets help find that excuse (maybe 23 year old male Corvette drivers who are Ph.D. candidates residing in a rural area are excellent risks). The initial "training set" for the neural net may be current underwriting practices. This would be the same starting point as a rules-based, inference engine driven expert system. Neural nets promise to gain the advantage over this because more factors can be considered and the "model" is self correcting based on actual sales results and loss experience. Andy Kowalczyk
F. Baube speculates that a neural net might hide an unlawful decision-process. Two comments on this: 1. A neural net -- indeed any other machine whether in black box or not -- cannot "be responsible" for actions that break the law. In Baube's example, it is the loan officer who makes the declaration that the loan is granted (or not), not the neural net. It is on the loan-officer's authority that the loan is granted; the machine has no authority at all. 2. It is easy enough to test whether a black box machine (e.g., a neural net) is breaking the law. One subjects it to a battery of test cases and sees what it is advising. In Baube's example, it would be satistically easy to determine if there is probable cause to believe the loan officer is engaging in redlining by relying on the advice of the machine. Such tests can be performed even if the "training set" has been lost. Peter Denning
Fred's comments will hold not only of neural nets but of any decision model trained from data (eg. Bayesian models, decision trees). It's just an instance of the old "gigo" phenomenon in statistical modeling. It is true that with neural nets is that they typically have so many parameters that it is difficult to "blame" a particular parameter setting for particular decisions. However, proper evaluation of complex statistical models cannot depend on examining particular parameter values anyway. Instead, evaluation requires the behavior of the model on held-out test data to be documented, and the test data to be made available for verification and certification purposes. If the model has acquired biases from its training data, those should be inferrable from its performance on its test data or new test sets, and the original developer should be held responsible for those biases. In the same way as ignorance of the law is not an excuse for breaking it, bad choice of training data is not an excuse. For example, if the data comes from the decisions of human beings (eg. loan officers) on particular cases, a lot of care must be exercised to ensure that those decisions are not affected by prejudice or confusion, rather than by examination of outcomes. On the other hand, if the data is derived from outcomes (eg., for loans, cases that ended in default vs cases that didn't, for loans with similar terms) and there is reasonable statistical coverage of all relevant case types, the model developer may be able to reasonably certify that any hidden correlations that the system discovers (eg. between location and default rates) are not demonstrations of intent to discriminate. Overall, the whole issue of evaluation, let alone certification and legal standing, of complex statistical models is still very much open. The traditional machinery of linear or log-linear statistics does not apply directly to nonlinear models such as neural nets or decision trees, so for example it is difficult in general to compute model error estimates, and thus the only source of such estimates is performance on test data. Unfortunately, these difficulties are often swept under the carpet in the hype that surrounds neural nets and allied methods. (This reminds me of a possibly apocryphal story of problems with biased data in neural net training. Some US defense contractor had supposedly trained a neural net to find tanks in scenes. The reported performance was excellent, with even camouflaged tanks mostly hidden in vegetation being spotted. However, when the net was tested on yet a new set of images supplied by the client, the net did not do better than chance. After an embarrassing investigation, it turned out that all the tank images in the original training and test sets had very different average intensity than the non-tank images, and thus the net had just learned to discriminate between two image intensity levels. Does anyone know if this actually happened, is is it just in the neural net "urban folklore"?) Fernando Pereira 2D-447, AT&T Bell Laboratories 600 Mountain Ave, PO Box 636 Murray Hill, NJ 07974-0636 email@example.com
F. Baube was concerned that a disingenuous bank might use a neural network trained to redline loans, and then destroy the training materials to hide the intent. Why would artificial neural networks have this advantage over the biological ones that banks have used to redline mortgages with for years? A human loan officer could be trained with redline-based materials which were later destroyed. Human neural nets are also opaque, although they often seem transparent. Tom Tom Janzen - firstname.lastname@example.org Real-time C, unix, VMS, AmigaDOS, Musical software, writer. See my video Dilettante at the DeCordova near Boston.
The decision-making process isn't really opaque. The are techniques for running a standard, feed-forward net "backwards", which will tell you what types of input are required in order to get the desired output. If you do this systematically, you can easily evaluate which areas of the input space result in a certain decision. Sure, if the training set contained redlining, the net would learn it - but that's part of the job, isn't it? Any other statistical technique (and in this case, a net trained by, e.g., backpropagation is just a non-parametric estimator) would suffer from the exact same problem. In both cases, the original data of which rules or weights are based could be hidden or destroyed. Jan Vorbr"uggen, Institut f. Neuroinforamtik, Ruhr-Universit"at Bochum email@example.com
But we already use a neural net for redlining -- the human brain which can hide all sorts of reasoning and, more likely, alternatives to reasoning. One of the big issues in red-lining and other areas where social policy meets business policy is that the argument is often over results and premises. More to the to the point the same issue occurs when business policy meets business policy. The opacity of neural nets reasoning is still an issue. Though, perhaps, more of an issue when trying to build systems with reproducible results. In general, however opacity can be a result of any complex system.
I remember a case similar to this coming to light in England about 7 years ago. It was in relation to a medical school that was using some form of artificial intelligence program to screen applicants. One item of data that was not available to the program was the race of the applicant, which should have helped ensure a relatively unbiased decision. However after complaints and some investigation it was determined that length of last name did have an effect on the outcome, and it was deemed that it was there because many students of, among others, Indian descent have significantly longer last names than the Anglo-Saxon applicants. So, it seems even this method can be caught, but obviously requires careful study to determine. John Turnbull
Please report problems with the web pages to the maintainer