On 6 May 2004, a Portland, Oregon, lawyer named Brandon Mayfield was arrested for his alleged involvement in the terrorist bombings of four commuter trains in Madrid. The attacks killed 191 people and injured 2000 others. But Mayfield had never been to Spain, and his passport at the time was expired. The sole evidence against him was a partial fingerprint found on a plastic bag in a van used by the bombers. The FBI’s Integrated Automated Fingerprint Identification System had identified Mayfield as a possible match, and three FBI fingerprint experts as well as an outside analyst confirmed the identification.
The analysts knew that Mayfield had converted to Islam, was married to an Egyptian woman, and had once represented a man in a child custody case who later turned out to be part of a jihadist group. That information swayed the FBI inquiry in Mayfield’s direction.
Spanish authorities, however, argued that the fingerprint belonged not to Mayfield but to an Algerian with a criminal record, Spanish residency, and terrorist links. They were right. It took almost three weeks from his arrest, but Mayfield was cleared of the charges and released from federal custody. The U.S. government eventually agreed to pay him US $2 million for the mistake and issued a formal apology.
Such high-profile cases grab the headlines and our attention, but they also point to an underlying problem with fingerprints—and with shoe prints, handwriting, and nearly every other form of classical forensics data. “The fact is that many forensic tests…have never been exposed to stringent scientific scrutiny,” a committee convened by the U.S. National Academy of Sciences concluded last year. One of the main problems with forensics evidence is that it must be analyzed and interpreted by a person, whose own theory of the crime can introduce a bias in the results. There can also be significant uncertainty in the analyst’s conclusions, but oftentimes that uncertainty is never quantified or conveyed to judges and juries.
And yet, these traditional forms of forensics evidence can be very helpful, provided they can be looked at objectively and the uncertainty of the results can be measured and properly explained. The relatively new field of computational forensics has sprung up to address those needs.
Computational forensics is not yet mainstream. But lots of research goes on in academic settings, such as at my own lab at the State University of New York at Buffalo, and eventually the courts may allow these techniques to be applied in criminal trials. Clearly, any method that can improve the analysis of evidence would be a good thing.
On the popular television series “C.S.I.” and its spin-offs, attractive forensics experts speedily uncover long trails of evidence that point unfailingly to the true villains. In the real world, crime scene investigations are rarely so cut and dried (and the people carrying them out are rarely so glamorous). More often, forensics experts must rely on a few murky clues, none of which is definitive in itself. Even then, as the Brandon Mayfield case revealed, the biases of the analysts can lead to erroneous conclusions.
To appreciate how computational methods could avoid such blunders, it’s helpful to understand how forensics data are used. In general, a forensics investigation attempts to match crime scene evidence with a known source. The first step is an analysis of the evidence, followed by a comparison of the evidence with data from known sources, and finally an independent verification of the results.