Month: April 2007

The more I learn from Alex Stoianov about the advantages of Biometric Encryption, the more I understand how dangerous the use of conventional biometric templates really is.  I had not understood that the templates were a reliable unique identifier reusable across databases and even across template schemes without a fresh biometric sample.  People have to be stark, raving mad to use conventional biometrics to improve the efficiency of a children's lunch line.

Alex begins by driving home how easy template matching across databases really is:

Yes, that’s true: conventional biometric templates can be easily correlated across databases. Most biometric matching algorithms work this way: a fresh biometric sample is acquired and processed; a fresh template is extracted from it; and this template is matched against previously enrolled template.

If the biometric templates are stored in the databases, you don’t need a fresh biometric sample for the offline match – the templates contain all the information required.

Moreover, this search is extremely fast, such as 1,000,000 matches per sec is easily available. In our example, it would take only 10 sec to search a database of 10,000,000 records (we may disregard for now the issue of false acceptance – the accuracy is constantly improving). Biometric industry is actively developing standards, so that very soon all the databases will have standardized templates, i.e. will become fully interoperable.

BE, on the other hand, operates in a “blind” mode and, therefore, is inherently a one-to-one algorithm. Our estimate of 11.5 days for just one search makes it infeasible at present to do data mining across BE databases. If the computational power grows according to Kim’s estimates, i.e. without saturation, then in 10 – 20 years the data mining may indeed become common.

Kim already suggested a solution – just make the BE matching process slower! In fact, the use of one-way slowdown functions (known in cryptography) for BE was considered before. The research in this area has not been active because this is not a top priority problem for BE at present. In the future, as long as the computer power grows, every time the user gets re-enrolled, the slower function will be applied to keep the matching time at the same level, for example, 1 sec.

Other points to consider:

• BE is primarily intended for use in a distributed environment, i.e. without central databases;
• the data mining between databases is even much easier with users’ names – you wouldn’t even need biometrics for that. We are basically talking about biometric anonymous databases – a non-existing application at present;
• if a BE database custodian obtains and retains a fresh biometric sample just to do data mining, it would be a violation of his own policy. In contrast, if you give away your templates in conventional biometrics, the custodian is technically free to do any offline search.

These arguments are beyond compelling, and I very much appreciate the time Alex and Ann have taken to explain the issues.

It's understandable that BE researchers would be concentrating on more challenging aspects of the problem, but I strongly support the idea of building in a “slowdown function” from day one.  The BE computations Alex describes lend themselves perfectly to parallel processing, so Moore's law will be operating in two, not one, dimensions.  Maybe this issue could be addressed directly in one of the prototypes.  For 1:1 applications it doesn't seem like reduced efficiency would be an issue. 

Why couldn't the complexity of the calculation be a tunable characteristic of the system – sort of like the number of hash iterations in password based encryption (PBE)?

Ann Cavoukian and Alex Stoianov have sent me a further explanation of the difference between the “glass slipper effect”, which seems to be a property of all biometric systems, and the much more sinister use of biometric templates as an identifying key.

Kim raises an interesting point, which we would like to address in greater detail:

“This is a step forward in terms of normal usage, but the technology still suffers from the “glass slipper” effect. A given individual's biometric will be capable of revealing a given key forever, while other people's biometrics won't.  So I don't see that it offers any advantage in preventing future mining of databases for biometric matches. Perhaps someone will explain what I'm missing.”

Let us consider a not-so-distant future scenario.  When the use of biometrics grows, an ordinary person will be enrolled in various biometrically controlled databases, such as travel documents, driver licenses, health care, access control, banking, shopping, etc. The current (i.e. conventional, non-BE) biometric systems can use the same biometric template for all of them. The template becomes the ultimate unique identifier of the person. This is where the biometric data mining comes into effect: the different databases, even if some of them are anonymous, may be linked together to create comprehensive personal profiles for all the users. To do this, no fresh biometric sample is even required. The linking of the databases can be done offline using template-to-template matching, in a very efficient one-to-many mode. The privacy implications explode at this point.

Contrast that to BE: it would be much more difficult, if not impossible, to engage in the linkage of biometric databases. BE does not allow a template-to-template matching — the tool commonly used in conventional biometrics. In each BE database, a user has different keys bound to his biometric. Those templates cannot be matched against each other. You need a real biometric sample to do so. Moreover, this matching is relatively slow and, therefore, highly inefficient in one-to-many mode. For example, running a single image against 10,000,000 records in just one BE database could take 0.1 sec x 10,000,000 = 1,000,000 sec = 11.5 days.

Kim is basically correct in stating that if an individual's real biometric image was somehow obtained, then this “glass slipper” could be used to search various databases for all the different PINs or keys that “fit” and, accordingly, construct a personal transaction profile of the individual concerned, using data mining techniques. But you would first have to obtain a “satisfactory” real image of the correct biometric and or multiple biometrics used to encrypt the PIN or key. All of the PINs or keys in the databases can and should be unique (the privacy in numbers argument) — as such, if an individual's actual biometric could somehow be accessed, only an ad hoc data mining search could be made, accessing only one entry (which would represent an individual privacy breach, not a breach of the entire database).

However, with BE, the actual biometric (or template derived from that biometric) is never stored – a record of it doesn’t exist. Without the actual biometric, data mining techniques would be useless because there would be no common template to use as one's search parameter. As mentioned, all the biometrically encrypted PINs or keys in the databases would be unique. Furthermore, access to the individual's biometric and associated transaction data would be far more difficult if a biometrically encrypted challenge/response method is employed.

In contrast, current biometric methods use a common (the same) biometric template for an individual’s transactions and, accordingly, can be used as the search parameter to construct personal profiles, without access to the real biometric. This presents both a privacy and security issue because not only could profiles be constructed on an ad hoc basis, but each template in a database can be used to construct profiles of multiple individuals without access to their real biometric. We thus believe that this alone makes biometric encryption far superior to standard current biometric methods.

Ann Cavoukian and Alex Stoianov

I had not understood that you can so easily correlate conventional biometric templates across databases.  I had thought the “fuzziness” of the problem would make it harder than it apparently is.  This raises even more red flags about the use of conventional biometrics.

Despite the calculation times given for BE matching, I'm still not totally over my concern about what I have called the glass slipper effect.  It would be a useful area of research to find ways of making the time necessary to calculate the BE match orders of magnitude longer than is currently the case.  If today it takes 11.5 days to search through 10,000,000 records, it will only take 4 hours in ten years.  By then the kids we've been talking about will be 16.  Won't that make it less than a minute by the time they are 26?  Or a quarter of a second when they're in their mid thirties?