D3.2: A study on PKI and biometrics

Title: PHYSIOLOGICAL (OR PASSIVE) METHODS

Fingerprint recognition

Fingerprinting has been used since 1901 to identify criminals in Great Britain and subsequently substituted the former Bertillon system which used certain skeletal patterns to identify a person. A year later fingerprinting was applied to analyse crime scenes by detecting fingerprints on objects touched by persons at the scene. About 1910 the first classification system developed by Edward Henry and was subsequently used by police forces and prison authorities throughout the English-speaking world. Since the mid 1980s automated fingerprint identification systems (AFIS) were tested and introduced in the USA and Australia. Following this development, fingerprinting was used more and more as an authentication method for accessing systems (access control, general use, i.e. verification). Today, fingerprint recognition systems are readily available e.g. for the PC where the sensor can be integrated into a mouse or keyboard.

Description of the method 

The fingerprint is not genetically defined, but develops randomly in the 3rd month of pregnancy on the tips of the fingers of the foetus [GEO01]. Fingerprinting uses these varying papillary structures on the tips of human fingers. Depending on the system used, seven to nine [AMB03] typical patterns can be differentiated. In addition to major fingerprint features shown in minor features such as the position of ridge ends and ridge bifurcations are used.

Figure 4‑: Seven common papillary patterns used by the FBI

For identification or verification of large numbers of persons, today automated fingerprint identification systems (AFIS) are used. These systems commonly use six steps for the analysis of fingerprints [BSI04], see :

Figure 4‑: Steps of analysis of fingerprints by AFIS

Step 1: Scan of the finger tips 

To take a picture of the fingerprint different types of sensors can be used: 

• Optical sensors (also used for latent fingerprints on a crime scene) 

• Electromagnetic field sensors 

• Polymeric thin film transistor sensors  

• Thermal sensors 

• Capacitive sensors 

• Pressure sensors 

• Ultrasonic sensors 

These sensors are used online and the picture of the finger tip is directly taken from the sensor. The traditional offline method of scanning a picture of a fingerprint is now used only in rare cases. 

Step 2: Optical image optimisation 

In this step the orientation of the picture and the optical quality (contrast etc.) is checked and if necessary adjusted automatically. If the result of this step is insufficient for further processing, a new scan is taken.  

Step 3: Preparation for classification and extraction of pattern 

In this step the papillary lines are extracted from the picture. 

Step 4: Classification of pattern 

This step is relevant for AFIS which is used particularly for crime investigation and forensic purposes. The fingerprint is assigned to one of three defined classes of fingerprints. 

Step 5: Extraction of identified pattern 

In this step the individual pattern such as ridge ends and ridge bifurcations and their position on the finger tip are identified using different algorithms.  

Step 6: Verification / Identification  

In this step the result of step 5 is checked against one (access control) or more (AFIS) stored templates. Determination of a class in step 4 can speed up the identification process because only templates belonging to the same class have to be checked against the sample. 

Value of the method

Depending on the sensors and algorithms used the equal error rate can be lower than 1%. In these cases the method can be more reliable than traditional photos e.g. in passports [BSI04]. This method performs very well compared with the seven criteria (seven pillars) used in [LIB05].

Limitations 

A major problem for verification purposes is spoofing of the sensors e.g. with faked fingerprints (see FIDIS deliverable 5.1). To prevent that, testing performed by advanced sensors can be used to detect if the sample is from a living subject [LIB05].  

Further limitations result from temporal or permanent changes of the fingerprints, leading to high failure to enrol (FER) rates. These changes may result from certain dermal diseases, injuries or hard labour (such as dirt, abrasion). Using good sensors and recognition software the FER can be less than 5% of the total population. If you look at people working in an office this rate reduces to less than 1%.

Privacy aspects 

Certain papillary patterns depend on the nutrition of the mother (and thus the foetus) in the 3rd month of the pregnancy [GEO01]. Various correlations between certain papillary patterns and corresponding diseases have been discussed over the years. Examples are leukaemia and breast cancer which seem to be statistically correlated with certain papillary pattern. A direct correlation in these cases is not known. Another example was published in 1986 where a correlation between certain papillary pattern and two forms of dementia was discovered. Since the 1950s various correlations between certain papillary pattern and human psychological characteristics and even homosexuality have been discussed.

Genetic fingerprinting, also called DNA profiling, was discovered and named in 1985 by Alec Jeffreys. It is a method to analyse and compare the DNA from various sources based on the observation that the DNA is unique for every person except for monozygotic (identical) twins. Today this method is applied for e.g., the genome of human beings, animals, bacteria, and plants. In the context of human beings, the method is mainly used to analyse the parentage (paternity tests) and for forensic purpose (e.g. investigation of crime scenes, accidents with many victims or genocide) to identify individuals and suspects from hair, blood, semen and other biological material [JOB04].

Applicable Parts of the DNA

Four genetic fingerprinting mitochondrial DNA molecules and the DNA molecules stored in chromosomes in the nucleus of the cells are used. Each DNA molecule contains four chemical units (called bases): 

• Adenine (A) 

• Guanine (G) 

• Cytosine (C) 

• Thymine (T) 

The bases are lined on a frame structure built of sugar molecules and phosphate groups. Position and sequence of the bases on this frame structure encode the genetic information. The DNA molecules are twisted into a double helix of two complementary strands (see ). In this structure each base pairs up with one other base: A with T and G with C. Genetic information is encoded in the sequences of the bases.

About 95% of the DNA of human beings do not appear to be coded (so called junk DNA) [AMB03], although new evidence has suggested that this hypothesis is not true. However, for genetic fingerprinting, the non-coded parts (parts carrying no genetic information) of the DNA are used. Notably, so called tandem repeats, which consist of repeats of the same sequence of bases, vary in the number of repetitions to a large extent among different people except for identical twins. Due to that variability they are of special interest when trying to discriminate individuals. The position of the repeats on the genome is called locus.

Currently, two types of tandem repeats are mainly being used for genetic fingerprinting: 

1. Tandem repeats, also called variable number of tandem repeats (VNTR) varying from core units of 2 to 103 bases and 9 to 100 repeats; there are thousands of loci known on each human chromosome,

2. Short tandem repeats (STR), varying from 1 to 30 repeats (so called alleles, see ). From 105 known loci of short tandem repeats (STR) about 20 currently are being used for genetic fingerprinting. STR are widely spread over the genome and some of them are neighboured to coded parts of the DNA [BEN01].

Analytical Procedures 

Currently, two methods for genetic fingerprinting are generally being used: 

1. The classical restriction fragment length polymorphism analysis (RFLP analysis), also called single locus probe (SLP) [JOB04], introduced by Jeffreys in 1985, using VNTR. A modern variant of this method is amplified fragment length polymorphism (AmpFLP or AFLP) [BYR01]. This method is largely replaced by variable number of short tandem repeats (VNSTR).

2. The newer method variable number of short tandem repeats (VNSTR or STRs), used since 1992 [BEN01].

The usual steps for both methods are as shown in and the result of the procedures is shown in (for technical details please refer to the Annex):

Figure 4‑: Analytical steps for Genetic Fingerprinting

Figure 4‑: Printout of a gel electrophoresis performed on three samples (1 to 3) and 4 loci (A to D)

Additional methods 

Among others commonly two subtypes of the VNSTR analysis are used [BEN01]: 

1. VNSTR analysis on the Y chromosome (Y chromosome STRs; e.g. locus 26-STR)

2. VNSTR analysis on mitochondrial DNA (mtDNA analysis, e.g. loci HVR1 and HVR2)

While the Y chromosome is passed through from father to son, the mtDNA is passed from the mother to all children and therefore can be used to establish maternal genealogical lineages. Furthermore mtDNA is quite stable and can particularly be used to analyse old or even fossil samples of DNA. 

In addition to the described methods, in forensic DNA analysis single nucleotide polymorphism (SNP) is increasingly being used [JOB04]. This method uses exchanged base pairs in the coded parts of the DNA. It is especially valuable when degraded DNA is being analysed. For identification purposes alone this method is of limited use because of the number of loci (around 50) that have to be analysed to get reliable results. Today this method is seen as an addition to VNSTR, not as a replacement.

Value of the Method 

The commonly used methods of genetic fingerprinting described are very reliable for the comparison of DNA samples and thus to identify persons and to clarify parentage. The reliability depends, among other aspects, on the number of analysed loci, the alleles found and the reference population used to determine the likeliness. 0.0000000004% of false positives (meaning that one out of 25 billion people have the same genetic fingerprint) are not unusual under good analytical circumstances. In many cases the reliability can be higher; results such as 1 out of 100 billion people are being discussed. But with respect to the number of parameters and their variable influence on the reliability (e.g. a found allele is, compared to the reference, very common or not) of the method no general statement on the reliability is possible [BEN01].

Technical Limitations 

Like every biochemical or chemical analysis, the quality of the results depends on the quality of the samples analysed (e.g. mixed samples with DNA from various persons for various reasons, contamination from bone marrow transplants, see [JOB04]) and the procedures carried out through the analysis. In Germany, one example is known where an error in the analytical procedures led to a false positive. These errors can be reduced simply by using good laboratory practice (GLP), A and B analysis in independent laboratories, working with external, independent observers and negative controls against staff involved in the analysis. There are numerous special procedures described to minimise problems arising from bad treatment of samples [FUC05]. But the quality of the sample taken from a certain habitat (e.g. a crime scene) remains a limiting factor to the quality of the analysis.

In the context of forensic analysis, depending on the analysed sample and their context, the methods in most cases only gives evidence that an identified person was at the crime scene. A positive identification does not tell directly, that the identified person committed the crime that was investigated. As with other biometric methods genetic fingerprinting can be spoofed e.g. by taking hair etc. from a person not involved to a crime scene. Further technological caused limitations of the methods are the time required to perform an analysis (about 24 hours for a genetic fingerprint) and costs. This limits the use of the method for verification purposes. But further development towards DNA-Chips or micro arrays could change that within a few years [BEN01], [SOK03]. 

Another limitation is that genetic fingerprinting as described does not work to discriminate identical twins. According to current data this applies to 1 of 250 persons in the world population [LIB05]. Further research to discriminate identical twins is currently being carried out. 

Privacy Aspects 

Among the possible consequences of the technical limitations discussed are:

• Genetic fingerprints taken for forensic purposes can in addition be used to determine the sex and parentage. 

• In some cases the ethnicity of the person behind the genetic fingerprint can be determined with some likelihood; this possibility was used to analyse the remains of multiple victims.

• Due to the neighbourhood of analysed STR to coded parts of the DNA there could be a relation between them and the analysed loci. This was discussed with the locus THO1 and the diabetes type 1 gene where the risk determined by analysis of a certain allele today is 0.12 % higher than the statistically average (0.4%) and thus is of limited practical value [BEN02]. It remains to be seen how this will be judged by commercial interests: insurance companies for example. Nevertheless, in Germany they are requesting to get access to the results of any performed genetic testing after 2011.

• Stored samples of DNA could be abused for other, privacy-invading forms of genetic analysis [JOB04]. 

• In addition ethical issues concerning the way genetic profiles are being stored in databases are being discussed [JOB04]. 

Some SNP analysis ethical issues including privacy issues are currently not sufficiently evaluated. Due to the use of coded parts of the DNA for this type of analysis, the correlation of genetically caused diseases with the base-pairs found at the SNP-loci can be expected in some cases.  

Face recognition

Face recognition is a sub-area of the general object recognition problem requiring the differentiation among objects only subtly varying from each other (i.e. faces) and thus presenting one of the most challenging computer vision problems. It is a non-intrusive biometric method dating back to the 1960s which has many commercial (for example entertainment, film processing, videophone, and teleconferencing), security, industry and law enforcement applications. Potential applications of an effective facial recognition system can be located in law enforcement for mug-shot identification, verification and access control for personal identification such as driving licences and credit cards, surveillance of crowd behaviour, anti-terrorism, as well as enhanced human computer interaction. And what is more, given the fact that humans are the centre of attention of many videos, face recognition systems offer a great number of advantages in areas such as content indexing and retrieval as well as video compression. Face recognition systems also provide additional security in human authentication systems which use other means of authentication, such as smart cards, passwords, etc. 

A recent application expected to spread widely concerns the reinforcement of security in mobile phones, PDAs and mobile devices in general through face authentication of the owner of the device. Taking into account the fact that the functionality of mobile devices has increased significantly and has been enriched with a variety of new services including personal information, such as address book, payment data and schedules, the need for protection of this information arises. This need for security combined with the fact that mobile devices are already camera equipped leads to the strengthening of the security of mobile devices by using a face authentication system. 

A facial recognition system may cover a wide range of areas, from well-controlled environments to uncontrolled ones, whereas they can be applied either to single images or to sequences of images, i.e. videos. In a controlled environment, frontal and profile images of human faces are captured with a uniform background and identical poses among the persons to be identified. However, the general face recognition task concerns an uncontrolled environment; significantly varying lighting conditions, occluded faces, varying facial expressions, faces appearing in different scales, poses and positions, complex and/or non-static background, complex foreground, more than one face contained in an image. Also, since facial features change through the years due to aging, the system should be robust enough to handle this requirement if the application requires it. 

Technical description of the methods 

The first task of the system is the detection of faces in the captured images. An image may contain zero, one or more faces. Given an image, the human detection and localisation module aims at automatically and reliably identifying and determining the position of all regions in the image which contain a human regardless of its three-dimensional position, orientation, pose and lighting conditions. Taking into account the fact that the complexity and efficiency of the face recognition step depends on the accuracy of the detection process in locating the human face in the image, the human face detection process can be regarded as very important. Any method used for human face detection/localisation should take into consideration not only the fact that faces are non-rigid and but also that they have a high degree of variability in size, shape, colour, texture, orientation, facial expression and occlusion. In case one or more faces exist in the processed image, then the location and size of each one should be reported to the system so that the marked area will be further processed in order to identify the detected face. 

Face detection constitutes a heavily researched area including a great number of developed face detection methods, which can be mainly classified into four categories: 

• Feature invariant approaches: which aim at detecting invariant facial features, in the sense that these features exist even when the viewpoint, orientation, or lighting conditions vary. The detected features are then used for face localisation. The basis of these methods is the observation that humans are able to easily detect faces in varying conditions and orientations, and thus features or properties invariant to these variable conditions and situations must exist. Examples of such features are the texture of the face (including hair and skin) and the facial colour. The latter is used on the strength of the results of scientific studies which indicate that the difference in the facial skin colour is mainly due to a difference in intensity rather in chrominance, and thus a number of facial skin colour models have been built in different colour spaces. In recent years a number of techniques that combine facial features in order to detect and locate faces have been presented. These features involve facial skin colour, shape and size for the determination of candidate areas in the image and then the verification phase takes place with the candidate areas being investigated for the existence of facial features such as nose, mouth, eye brows and hair.

• Knowledge-based methods: which are based on the effort to encode human knowledge of what a typical face looks like. Most often this knowledge is composed by rules which capture the relationships between facial features, regarding their relative distances and positions. Examples of such rules can be the fact that a face includes two eyes which are symmetrical to each other, a nose and a mouth. However there remains the danger that if the rules are strict, then they may fail to detect faces that they do not satisfy all the rules, whereas if they are too general, they may provide too many false positives. Moreover, the detection of faces in different poses becomes a challenging task since all possible cases must be taken into account. Nevertheless, as far as frontal faces are concerned, these methods can efficiently detect them in uncluttered scenes.

• Appearance-based methods: the models which are learnt from a large set of training images of faces of variable features with the learned facial characteristics being stored in the form of either discriminant functions or distribution models. These models are used for face detection by being applied to the images to be examined.

• Template matching methods: which involve the matching of an input image with standard patterns of a face describing either the face as a whole or the facial features separately. This matching involves the computation of their correlation values and thus the existence of a face is determined by these values. The drawbacks of these methods mainly lie in their inability to provide reliable results in an efficient way in.

The Eigenface approach, one of today’s most advanced methods for face recognition, is described in detail in the annex 2. 

Value and (technical) limitations 

The Eigenface recognition method has several advantages, such as the fact that no significant low-level or mid-level processing is required before raw intensity data are used for learning and recognition, there are no requirements on knowledge of geometry and reflectance of faces and it is simple and efficient. However the results of the implementation of this method reveal some limitations, such as that its efficiency and accuracy decrease under varying pose and illumination, learning is time consuming. The results concerning feature-based face recognition show good accuracy, with recognition rates of up to 90%. 

The use of neural networks in these areas is rather popular achieving high accuracy – varying from 90% to 99% depending mainly on the number and variations of trained samples. However, their use during the feature extraction phase can be rather time consuming due to their complexity and thus are usually used in the pattern recognition phase. The nearest neighbour classifier is rather fast –faster than neural networks especially concerning the training time. Nevertheless, as far as accuracy is concerned, results have shown that a neural network is able to achieve the same accuracy with the nearest neighbour classifier using far less eigenfaces, and a higher accuracy when using the same number of eigenfaces.  

Privacy aspects 

The main question arising as far as the use of face recognition systems in public areas is concerned is whether it invades privacy. Taking into account the fact that these systems operate based on physical human characteristics that are constantly exposed to the public, many people regard that there stands no matter of privacy invasion as far as the access to images of facial features is concerned. However, there should be notification to people entering a public area where camera monitoring takes place, so that they are given the opportunity to make a choice whether or not they wish to be included in the captured videos. This choice of course is rather fictitious in public areas such airports, however these areas are already video monitored. 

An issue concerning face recognition systems used for surveillance and security reasons is what level of official procedure is required for a person’s image to be entered into the image database used by the face recognition video surveillance system. In the case of an arrest warrant having been issued, then things are rather simple. Nevertheless, there remains the danger of abuse of such systems from people simply wanting to track the life of other people. The system for instance is technically able to capture the image of every person entering the area covered by cameras and store it for future use. The only limitation could be the size of the image database. One solution to this could be the maintenance of a record including: the entrance date of the image into the database, the reason for this action, as well as whose request triggered this action. Moreover, there is the possibility that the image databases and the face recognition surveillance systems might be networked together, so that in the case of detection of a particular person in an area (e.g. airport), then similar systems in other areas are notified (for example railway stations, ports and subways). Such a scenario could lead to the unstoppable monitoring of a person’s entire movements. 

Another aspect of face recognition is that from a picture a specialist can diagnose various diseases such as marfan-syndrome, down-syndrome or crouzon-syndrome and acne [KRA05].