Improving cyber security using realistic synthetic face generation

Lead Research Organisation: University of Kent
Department Name: Sch of Physical Sciences

Abstract

The grant application outlines a novel programme of research that questions the uniqueness of facial identity and investigates the use of computer generated face imagery in the area of cyber security.The popularity of the human face as a biometric remains strong despite the introduction of many competing modalities. People are accustomed to being identified by their facial appearance whereas other biometrics such as fingerprints and iris recognition feel more invasive. A programme of research that investigates the concept of identity that is highly relevant to cyber security is proposed. In addition we will develop a novel cyber security application based on facial identity and evaluate its practical security level. Work of this nature has relevance beyond the scope of the project. For example, border control officers routinely verify a person's identity using passport photos but what is the fundamental limit on the ability to achieve this task reliably?

Planned Impact

The proposed project will provide impact through: its deliverables, potential for future product development, engagement of the general public through interactive online experiments and by advancing scientific understanding in the area of cyber security.
A deliverable of WP1 is a MATLAB toolbox for synthesising face images. The toolbox will comprise a face land marking tool, code for generating the PCA/GMM model and a user interface that allows batch generation of lifelike face images. We anticipate that the toolbox will be highly beneficial to researchers working in the areas of cyber security, biometrics, computer vision and psychology. The toolbox will be made available for download, free of charge, from the MATLAB File Exchange.
The source code for the prototype cyber security application, developed in WP4, will be uploaded to the collaborative software development site www.github.com to encourage the involvement of developers in the wider security community.
The proposed programme of research will result in Gold Standard Open Access publications and be disseminated at prestigious conferences such as ACM Conference on Computer Communications Security (CCS), which is the highest ranked conference in Computer Security in the World.

Publications

10 25 50
publication icon
Alsufyani A (2019) Breakthrough percepts of famous faces. in Psychophysiology

publication icon
Liu J (2019) Dynamic spectrum matching with one-shot learning in Chemometrics and Intelligent Laboratory Systems

publication icon
Osadchy M (2017) No Bot Expects the DeepCAPTCHA! Introducing Immutable Adversarial Examples, With Applications to CAPTCHA Generation in IEEE Transactions on Information Forensics and Security

 
Description The project explored three main themes within the area cyber security. (1) The first of these improved the capability of CAPTCHAs, a tool that permits humans to access web based resources but prevents bots (machines) from doing so. Prior to our work (output published in IEEE TIFS, 2017, 24 citations) many researchers believed that CAPTCHA systems had become ineffective due to advances in classification accuracy afforded by convolutional neural networks (CNN). Contrary to this view point our work utilises adversarial examples that have been shown to fool neural network classifiers for recognition tasks that are easily solved by humans. (2) We have demonstrated that there are fundamental limits on the number of distinct identities that can be recognised by neural network face classifiers and have estimated the entropy of such systems. The entropy is lower than expected which has security implications for the future face recognition. (3) In the third area of work we propose a new concept in biometric template protection in which genuine facial identities are hidden within a data set of computer generated face images.The proposed 'Multi-face' system protects users' privacy and makes exfiltration of a face database infeasible and identifiable.

The collaborators continued to work together on new projects beyond the programme of funded research. This has resulted in a further two publications in the area of chemometrics. Additional funding (following the initial grant) supported a pilot study leading to a larger industrial project (funded by Innovate UK) Forensic Facial Identification using the Fringe P3 brain wave response (EEG-FIT).
Exploitation Route The work has implications for developers of face recognition systems and protection of online services against bots.
Sectors Chemicals,Digital/Communication/Information Technologies (including Software),Security and Diplomacy

 
Description BEIS additional funding
Amount £90,000 (GBP)
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Public
Country United Kingdom
Start 01/2018 
End 03/2018
 
Description Forensic Facial Identification using the Fringe P3 brain wave response (EEG-FIT)
Amount £309,057 (GBP)
Funding ID 104590 
Organisation Innovate UK 
Sector Public
Country United Kingdom
Start 02/2019 
End 07/2020
 
Description Facial composite construction using brainwaves 
Organisation VisionMetric Ltd
Country United Kingdom 
Sector Private 
PI Contribution The additional funding provided by BEIS/EPSRC funded a pilot study to investigate the use of EEG in facial identification. We subsequently applied for, and won, funding (from Innovate UK) to support the development of a commercial facial identification system based on this idea. Funding for the project 'Forensic Facial Identification using the Fringe P3 brain wave response (EEG-FIT)' was awarded to University of Kent spin-out company, Visionmetric (£216,832) and School of Computing, University of Kent (£92,225). The 18 month project started in Feb 2019.
Collaborator Contribution Visionmetric - facial identification software development and know-how. University of Kent - EEG facilities and expertise.
Impact Anticipated EEG based facial identification system.
Start Year 2018
 
Description Improving cyber-security 
Organisation University of Haifa
Department Department of Computer Science
Country Israel 
Sector Academic/University 
PI Contribution Developing software tools and face stimuli for experimental work, experimental design, running tests using a computer cluster, coauthoring manuscripts.
Collaborator Contribution Experimental design, implementing machine learning architectures, coauthoring manuscripts.
Impact N/A
Start Year 2015
 
Description Spectrum matching using deep neural networks 
Organisation University of Haifa
Department Department of Computer Science
Country Israel 
Sector Academic/University 
PI Contribution Expertise in chemometrics and Raman spectroscopy, and access to associated instrumentation. Co-authoring of manuscripts.
Collaborator Contribution Dr Rita Osadchy (Haifa) Expertise in the area of deep learning. Co-authoring of manuscripts. Jinchao Liu (Visionmetric) Design and implementation of deep learning architecture. Co-authoring of manuscripts.
Impact 2x publications (see publication list). Contributed to expertise
Start Year 2016
 
Description Spectrum matching using deep neural networks 
Organisation VisionMetric Ltd
Country United Kingdom 
Sector Private 
PI Contribution Expertise in chemometrics and Raman spectroscopy, and access to associated instrumentation. Co-authoring of manuscripts.
Collaborator Contribution Dr Rita Osadchy (Haifa) Expertise in the area of deep learning. Co-authoring of manuscripts. Jinchao Liu (Visionmetric) Design and implementation of deep learning architecture. Co-authoring of manuscripts.
Impact 2x publications (see publication list). Contributed to expertise
Start Year 2016