ICIAC 2011 Seminar stream 5B

Countering terrorism with technology (I)

Abstracts and slides

Enhancing border security with automatic emotion recognition

Les Ball, Lecturer, University of Abertay, Dundee

Organised terrorist cells continue to pose a real threat to national security, and border controls such as airports and seaports remain at the centre to provide rigorous and robust processes of person identification. Some major terminals, as a consequence, now acquire the processing of biometric data such as fingerprints and facial snapshots of individuals. The process overall tends to be time consuming and cumbersome due to the intrusive nature of data acquisition and poor human computer interaction design. Moreover, dependency on a single mode of data acquisition for biometric identification is less favourable than a multi-modal approach. Emotion recognition is another aspect of surveillance which is to date relatively underexplored. This aspect has the potential to enrich the biometric process by acquiring additional information on “mood” or “state” of an individual. The question then becomes not only “who is this?” but also “is this person’s behaviour suspicious?”. Biometrics are classified as physiological (e.g. face, fingerprints) or behavioural (e.g. voice, gait) characteristics. The additional aspect of emotion recognition gives rise to the possibility of enriching and enhancing the decision support inherent to surveillance operations. The use of emotion recognition technology has recently been brought under the spotlight in terms of its potential to support the National Security agenda (Cooper, 2010). We have already presented our biometric work on keystroke dynamics at the Crime Science Conference and are currently working with a local Police Force that has expressed an interest in the use of this system in one of their incident rooms. Here, we propose extending this work to develop a biometrics portal that incorporates multi-modal fusion in the identification of individuals’ behaviour from surveillance cameras and other appropriate sources. The research would therefore contribute to a most suitable portfolio of measurements to best predict the emotional state and identity of an individual. In a biometric system, the quality of the system can be assessed by enrolling valid users and then testing the system against those enrolled and those posing as impostors. A standard psychological methodology for the assessment and elicitation of emotional states (e.g. see Coan & Allen, 2007) will be applied during biometric data collection. Based on the findings of our earlier study, about 35 participants should be adequate for creating a sufficiently large database for biometric analysis. The participants will be enrolled into the system using various identified biometric characteristics, such as face, voice, hand geometry and gait Further analysis will be applied on how each one may link to human ‘affect’ using the standard emotion elicitation procedures. Our resulting tool will be implemented to not only identify but also display any affective cue exhibited using the participants and impostors. The appropriateness of which characteristics to collect is also one of the key research objectives, as the intention is that the development of a biometrics portal should operate as unobtrusively as possible. Taking DNA samples, for example, would be inappropriate, while pupil dilation could perhaps be included. The key question of the research is to design, implement and evaluate a physical portal for multi-modal biometric data capture and analysis to recognise emotional content in human subjects and support decision-making in real-time security environments.

Presenter's slides: ICIAC11_5B_LBall

Text analytics for detecting terrorist activities

Claire Brierley, Senior Research Fellow, School of Computing, University of Leeds

The EPSRC-funded Making Sense project [1] aims to create an interactive, visualisation-based decision support assistant for intelligence analysts across a range of applications. It constitutes novel technology in that system design and development is informed by psychological findings about the behaviour of analysts, and system outputs are interactive visualisations of associations and relationships in multi-modal data which facilitate sense-making and human reasoning: the new field of Visual Analytics [2]. The modular systems architecture (data collection ® fusion and inference ® analysis ® visualisation) includes a text extraction module where the challenge is twofold: summarising content in high volumes of noisy text (e.g. non-standard, incomplete and unreliably-sourced data); and discovering links between documents by deriving genre-specific, linguistic characteristics of ‘suspicious’ documents.

We compare automated Text Analytics techniques for extracting text-based content and find that Information Extraction is the predominant technology in commercial intelligence analysis software. However, this approach may constrain type of information gleaned from text, focussing on information for populating pre-determined event scenario templates [3]; it also pre-supposes well-formed text. We also review datasets used in intelligence and security research and find that common problems are lack of authenticity and lack of truth-marked data for algorithm development and benchmarking: machine learning relies on truth-marked data for training and evaluation of learning schemes and usefulness of feature sets.

Text Analytics presupposes a corpus or sample of language texts capturing empirical data on the behaviour being studied; it then applies computational techniques for quantitative, empirical analysis and modelling. Our approach to text extraction implements Keyword Extraction: a statistical process for retrieving keywords via formal comparison of word frequency distributions in a test set relative to their expected frequencies in a suitable reference set. Keywords (and phrases) are thus significant by virtue of their uncommon frequency or infrequency. Having derived such items from truth-marked data, we then test their efficacy as classificatory features for retrieving like texts.

To solve the problem of data availability, we simulate the task of identifying similar texts of interest in an open-source document collection which has been analysed and classified by academic scholars. Using the Qurany ontology browser [4], we isolate a subset of documents associated with a target concept (not suspiciousness, but something analogous in scope), where the association is determined by domain experts (i.e. scholarly opinion informing the ontology itself). We then evaluate various feature sets and classifiers for the text extraction task, extending to discovery of similar material in a different collection.

We have used Keyword Extraction in simulated scenarios to gist the content of terror-related news reports[5], and to help determine the spread of a viral epidemic [6], and we are keen to apply this technique to a corpus compiled from the volumes of data collected in evidence about the recent riots in Britain (e.g. interview and mobile phone transcripts etc). The EPSRC-funded Making Sense project has CPNI as its main stake-holder; we hope to work with Police and other agencies in further research and development of Text Analytics for Crime and Intelligence Analysis.

Presenter's slides: ICIAC11_5B_CBrierley

Page last modified on 24 nov 11 15:25