Get to know more about ICMedia 2012 Presentations:
Title: Collection, organization, classification and analysis of digital evidence in present days: a perspective
Dr. Anderson Rocha (Campinas University)
Abstract: On this lecture I will discuss topics related to Computer Forensics and Forensic Document Analyses. Some of these topics concern the collection, organization, classification and analysis of digital evidence using techniques of machine learning and computer vision. Also I will present the emerging research field called forensic analysis of digital documents (e.g., images and videos) to verify its authenticity and integrity. Basically, I will focus on four sub-areas of research: camera identification, the distinction between natural images and computer generated images, the identification of tampering in digital documents and the determination of operating history suffered by a digital object and its relationship to other related objects (digital phylogeny). With regard to identification of the origin of an image or video, the goal is to identify a particular model of camera or camcorder, or the exact camera used to capture an particular image or video classified as potential evidence. The purpose of the distinction between natural images and computer generated images is to point out features that might allow the identification of synthetic images when compared to natural ones (not computer generated). The purpose of detecting tampering of documents is to establish its authenticity, or to expose any kind of tampering. Finally, the goal of digital phylogeny is the determination of the modifications structure suffered by a set of digital objects (e.g., images and videos) and the construction of the phylogenetic tree relating these elements. Whenever possible, I will emphasize the results obtained by our research group in IC/Unicamp in a way that our research can be put into practice by forensic experts.
Title: New trends in forensic evidence evaluation - The place of speech science in an ongoing paradigm shift.
Dr. Anders Eriksson (University of Gothenburg)
Abstract: Forensic evidence evaluation, and in particular the presentation of such evidence in court has been the topic of a lively debate over more than a decade now. Among the problems that have been identified are:
Disparities in the Forensic Science Community
Lack of mandatory standardization
Lacking measures of performance
Lack of certification and accreditation
In the two surveys cited in the references (Law Commission, 2011; National Research Council, 2009), forensic speech science is not studied or even mentioned, but most of what is said is nevertheless highly relevant also for speech science evidence. I will begin by describing the situation in Europe and Australia based on an informal survey I have recently conducted and also illustrate some of the questions and problems with examples from actual case work. I will also illustrate, by example, the problems that may arise when different types of evidence that may be difficult to compare come into conflict and the difficult task of making the concept of Likelihood Ratio comprehensible to a jury. I will also talk about the problematic nature of the concept of individualization which is, directly or implicitly, underlying many types of forensic evidence presentations.
Title: Forensic Speaker Identification considering Automatic Voice Comparison Systems
Dr. Isolde Wagner (Federal Criminal Police Office, Germany – BKA)
Abstract: During recent years, several automatic voice comparison systems have been developed for the application in forensic speaker identification. In order to account for validity and adequate performance estimation, the Federal Criminal Police Office of Germany (Bundeskriminalamt - BKA) developed its own system called SPES, which relies on authentic forensic case databases. Research at the BKA has shown that authentic forensic case data cause performance reduction compared to laboratory speech. This mainly is a result of the variability in case recordings caused by different recording channels, durations, and speaking styles. Hence, forensic phonetic experts have to account for these conditions when applying an automatic system in forensic voice comparison casework.
Title: Likelihood ratios for forensic speech science in the UK: the state of play in 2012
Erica Gold, University of York
Abstract: Numerical and verbal likelihood ratios (LRs) are used by approximately 20% of experts worldwide in forensic speaker comparison casework. A likelihood ratio is an estimation of strength-of-evidence based on the probability of occurrence given the prosecution hypothesis divided by the probability of the occurrence given the defense hypothesis. A vast majority of experts implementing LRs are practitioners who incorporate an automatic speaker recognition system (ASR) into their analysis. The automatic system more readily allows for the presentation of LRs as it is one of its easily selectable options for conclusions (Gold and French 2011). However, there are many factors to take into account for the practical implementation of LRs (as outlined in French et al. 2010), the most prominent being the lack of population statistics which are essential to the calculation of any numerical LR. To remedy this, a large amount of research has started in an effort to collect population statistics. Essential to this endeavour is the collection of population statistics from a large homogenous database. Many researchers in the UK have recently been relying on a forensically relevant database, the Dynamic Sources of Variability in Speech (DyVis) (Nolan 2009), which consists of 100 male speakers ranging in age from 18-25, all of whom are speakers of Standard Southern British English (SSBE) for collecting population statistics.
The vast majority of LR research that has been conducted in forensic speech science has focused on vowel formants (Zhang, Morrison, and Thiruvaran 2011, Morrison 2008, Rose 2006, Loakes 2006, Kinoshita 2001). However, there are a large number of other speech parameters that are considered discriminant by experts for use in forensic speaker comparison casework (i.e. voice quality, VOT, speaking tempo, non-linguistic features, consonants) (Gold and French 2011). Therefore, it is vital to make an effort to incorporate such features into LRs alongside vowels. There are, however, many factors that prove challenging to such a task, including the discrete and qualitative nature of certain features, the high within speaker variation that can be found for some parameters, and the lack of population statistics for additional accents and features.
Challenges facing the complete implementation of numerical LRs into forensic speaker comparisons will be discussed in relation to research conducted using the DyVis database to analyze click production (Gold 2012a) and articulation rate (Gold 2012), as well as population statistics being collected by other researchers at the University of York. Additional challenges such as feature interdependencies (Gold and Hughes 2012) and effects of variability in the reference data (Hughes 2011) will also be discussed in relation to research conducted using the DyVis database. A summary of this research will also be given in efforts to provide a realistic overview of the position of LRs for forensic speech science in the UK in present day.
Title: 3D video analysis at the Netherlands Forensic Institute.
Dr. Jurrien Bijhold (Netherlands Forensic Institute)
Abstract: In some investigations, like large scale accidents or terrorist attacks, it is required to analyze the movements of cars and pedestrians from large amounts of video footage from public and private surveillance camera's and hand held cameras like phone cams. For getting the capabilities to perform such a task, the Netherlands Forensic Institute has carried out a project for 3d video analysis in cooperation with the Netherlands Police, the University of Amsterdam and the National Institute of Applied sciences. In this project, methods and procedures have been developed for collecting, downloading and synchronizing video footage, and analysis using 2d and 3d maps of cities. The methods also allow for the introduction of intelligence from witnesses and cell phone communication data. As an introduction to this project, this presentation will start with an overview of the work done by the group for image analysis at the Netherlands Forensic Institute. This group is active in the fields of video technology, image authentication, file repair, camera identification, photogrammetry, facial image comparison and forensic applications of 3d modeling. Extra attention will be given to the cases and experiments that lead to the start of the project for 3d video analysis. After this introduction, experiments and cases are presented that were carried out during the project for testing the new methods and procedures. All results from this work, including the results from the scientific research on methods for tracking and tracing, and fast 3d modeling , will be presented and discussed. The presentation ends with an overview of the possibilities for 3d modeling and 3d video analysis that are presently available in the Netherlands.
Title: Challenges in image and video forensics
Dr. Zeno Geradts (Netherlands Forensic Institute)
Abstract: In this presentation a short overview will be given of developments in image and video forensics. The state of the art methods on digital video and image authenticity will be presented, also with several examples. New developments in the field of camera identification with Photo Response Non Uniformity will be discussed, as well as possibilities to use the methods on larger databases, or on on line databases for instance YouTube. Furthermore the techniques for facial comparison also in relation to photogrammetry for height measurement will be discussed. Also methods for comparison of other biometrics such as hands will be included. New R&D in the field concerning such as detection of heart rate from faces and possibilities for video authenticity and its use in forensic science.
Title: A Snapshot of Security and Privacy in Biometrics
Dr. Walter Schereir (University of Colorado)
Abstract: While we might consider biometrics to be just another useful type of security technology, the concept of personal identity is important from several perspectives. From a cultural perspective, the more the world converges, the more individual cultures wish to maintain their separate identities. From an individual perspective, the greater the population and the tendency to reduce people to stereotypes, the greater the desire to establish an individual identity. There is, however, another level where identity and the verification of identity is becoming increasingly important in relation to all manner of transactions, from those related to mobility, to those related to legal, and political, rights and obligations, to those of financial and economical transactions. The intrusion of technology into these areas is not new, but their heightened visibility and ubiquity can create anxiety. This holds particularly true for biometrics, where social acceptance will be key to its long-term success. In this talk, the social problems of biometrics will be considered from a security and privacy perspective, and the research area of “biometric template protection” will be introduced. This exciting new field within biometrics incorporates ideas from computer vision, pattern recognition, cryptography and network security, making it a truly remarkable cross-disciplinary endeavor within computer science. When pattern recognition and cryptography are combined, a new class of algorithms and protocols enables privacy preserving identity verification, including a full “Biocryptographic Key Infrastructure” with applications to the Internet and beyond. This talk will briefly survey critical technologies in this emerging area, with an emphasis on their potential for surveillance, access control and remote authentication for events such as the World Cup and Olympics.
