Project description

Project Overview

With various forms of biometric technologies becoming available, there is a growing need for scientists who are able to assess the merits of these technologies when applied to forensics. The Marie Curie ITN "Bayesian Biometrics for Forensics", or BBfor2, will provide a training infrastructure that will educate Early Stage Researchers in the core biometric technologies of speaker, face and fingerprint recognition, as well as the forensic aspects of these technologies. According to modern interpretation of evidence in court, biometric evidence must be presented as likelihood ratios. The calibration of likelihood ratios of individual behavioural and physical biometrics and of combinations of biometric modalities, including measures of the quality of the traces, is a unifying topic in all research projects in this Network.

The Network will conduct research in the core biometric detection areas of speaker, fingerprint and face recognition. The working experience in forensic research environments will educate the ESRs to become experts in the aforementioned biometrics modalities, with the application of forensic research in mind. The multi-disciplinary training makes the ESRs very employable in forensic sciences and governmental bodies such as law enforcement institutions, as well as research and development departments of industry.

The training of ESRs will be realized as individual PhD projects at various research labs, including a forensic institute. Apart from training at their host institute and secondments with other network partners, the ESRs will receive training in dedicated Summer Schools on Biometric Signal Processing, Bayesian Techniques in Forensic Applications and Legal Issues in Forensic Applications. The Network combines 8 European Universities and a leading Forensic Institute; it is augmented by a biometric industrial, a research institute and a forensic science services provider, where secondments of the ESRs will take place.

Concept and Project Description

Biometrics have always played an important role in forensic research. Perhaps the most well-known example is the traditional fingerprint, the utilization of the characteristic line patterns that the skin of human fingers shows, which are assumed to be unique to every human being. In forensics, we may define two separate use cases for biometrics, namely investigation using a trace to narrow down the search for potential suspects, and evidence using a trace as support to convince the court (the judges or the jury) that the trace is produced by a particular suspect.

The common belief of the indexing capabilities of fingerprints, pinpointing a single fingerprint unambiguously to a unique person, is so wide-spread that a similar terminology is sometimes being used in other biometric fields, e.g., the misleading term "voiceprint" for speaker recognition. However, from the field of speaker recognition we have learned to apply the Bayesian framework of probability, and we know that it is not possible to identify a person from its voice. There is not a way in which we can compute the posterior probability (or posterior odds) that a trace is produced by the suspect, given the trace as data, by only applying the biometric. Rather, the theory tells us, we can only compute a likelihood-ratio, the ratio of the likelihood that the trace is found assuming it is produced by a particular speaker, versus the assumption it is produced by somebody else.

The Bayesian framework of interpreting biometric technology has several advantages. Firstly, it brings the two use cases in forensics closer together. In fact, the same technology can be used for the purpose of investigation and evidence, if the biometric can be shown to produce calibrated likelihoodratios. This can potentially make the investigation more effective as traces that help in finding a suspect of a crime can also serve as evidence. Secondly, we can use similar methodologies for the various biometrics for determining the calibrated likelihood ratio. Thirdly, we can combine multiple traces in different biometric modalities in the same framework.

The Project Objective is to develop a Bayesian framework of presenting biometric evidence in a way that is consistent over the various biometric technologies and that takes into account the circumstances encountered in the collection of forensic biometric traces.

This Objective will be achieved by recruiting Early Stage Researchers that will focus on the various aspects involved, by conducting research that will lead to a doctoral degree at the end of their training. The aspect of consistency will be achieved by the organisation of workshops and summer schools every half year.

Scientific and technological objectives of the research and training programme

Research: A characteristic challenge raised by forensic application of biometric technology is that the trace , unintentionally left behind by the perpetrator, is produced in uncontrolled conditions. This leads to a wide variability in sample quality, and hence to a lower performance of the biometric detector. Related to the typical usage of biometric systems for authentication is the issue that some industrial systems attempt to do matching of a sample with enrolled identities, as happens in, e.g., fingerprint recognition. For investigation purposes in the forensic domain it is useful to obtain similarity scores from the biometric detector, however, for a system that only find a match are systems that do identification . For forensic application, there is very little use in an identification system, as a matter of fact, there is very little application of identification systems in general.

A second issue that is of importance is calibration . A basic biometric technology will provide a similarity score between trace (test sample) and model (training sample). When consistently high scores are obtained when test and train identity are the same, and low scores when they are different, the biometric system is said to be a good detector . As such, it can be used in investigation, to filter unlikely identities out from more promising suspects. The absolute similarity score is not so important in this application. If there is deterioration of the trace (for instance, a bad quality telephone line in a wire tapped conversation) similarity scores may all become smaller. But as long as the similarity score of the perpetrator is relatively high, it will be found as a potential suspect during a search. However, if such a trace is to be used as evidence in court, a mere similarity score no longer suffices. In that case, it is required to present a value of the likelihood ratio that has an absolute interpretation, and not a similarity score that only has an interpretation relative to other similarity scores. The process of making a similarity score to be interpretable as a likelihood ratio is called calibration. Calibration of a basic biometric detector is not a trivial problem.

Related to the calibration problem is the use of a reference database for the suspect. The denominator of the likelihood ratio is the likelihood that the trace is found assuming somebody else than the suspect. For a likelihood ratio to be well-calibrated, this denominator must be estimated properly. In the biometric calibration, this denominatormust be represented by "others" since the true perpetrator will not be known to the forensic specialist if the defense hypothesis is true. How to choose such reference database is a difficult issue, and will have different aspect for the different biometric modalities. Given these issues that exist in the forensic application of biometric technologies, we can state the scientific objectives of BBfor2:

  • to make biometric technologies applicable in the forensic domain for the purpose of investigation
  • to develop a framework to make biometric technologies represent evidence in a way that can be presented in court.
The research is carried out by 15 ESRs and ERs who will each conduct research in an individual research project, which are grouped into larger work package.

Training: The twelve ESRs will all be working 36 person-months in BBfor2, and additionally, three Experienced Researchers (ER)s will work 24 months each, consecutively. Two ERsite to guarantee cohesion between the various research projects and continuity within the Network.

The network as a whole undertakes to provide a minimum of 504 person-months of Early Stage and Experienced Researchers whose appointment will be financed by the contract. Quantitative progress on this, with reference to the table contained in Part C and in conformance with relevant contractual provisions, will be regularly monitored at the consortium level.

The training will be organized in 4 work packages organized by topic: Speaker Recognition, Face Recognition, Fingerprint recognition and Overall aspects. Within each research topic, 2-4 ESRs will carry out the research, supervised by two members of the BBfor2 Network from different partners. The research will be carried out as employee of the organisation of the First Supervisor.