Cancelable Iris Biometrics using Block Re-mapping and Image Warping Andreas Uhl Department of Computer Sciences University of Salzburg, Austria uhl@cosy.sbg.ac.at http://www.wavelab.at/
Outline • Introduction & Motivation • Iris Recognition • Cancelable Iris Templates: Block Re-mapping and Warping • Experiments – Impact on Recognition Performance – Key Sensitivity • Conclusion Andreas Uhl 1
Motivation Revocability problem : The use of biometrics comes with different problems as compared to conventional authentication systems. As biometric features are specific to an individual person, they cannot be changed (or not often, one person for example has only ten fingerprints and two iris patterns available). So where a password can simply be changed or an e-card invalidated, this is not possible with biometrics. Application-specific keys : In the same way, it is not possible to use different keys for different applications - for example if one wants to use a different key for the bank account and for access to the workplace computer. Privacy problem : some biometric modalities are not even very secret - for example low-quality fingerprints are left everywhere, and eye images could be captured by hidden cameras. This does not only open the possibility to forge biometrics (like showing a picture of a person to a face recognition system), but also can raise privacy concerns. Large databases with biometric data would gain the potential of misuse, for example by cross-matching. Andreas Uhl 2
Cancelable Biometrics Security is implemented by applying a key-dependent transformation to the captured biometric signals or templates. The transformation must be non- invertible so that the original data cannot be reconstructed from the stored transformed version. At the same time matching still has to be possible with the distorted version. For iris recognition, the following techniques have been suggested: • Chong et al.: a binary iris code is generated by thresholding inner products of the feature vector with randomly generated vectors. • Chong et al.: iris pattern images are multiplied with a random kernel in the frequency domain, then MACE filters are created using several training images providing a similarity measure for each test image after multiplying it with the user-specific random kernel as well. • Zou et al.: employ circular shift to polar iris image rows or iris codes plus addiitonal combination of selected rows into new rows; alternatively, they suggest to use salting , which combines a random pattern with the iris texture or the iris code. Andreas Uhl 3
Iris Texture Extraction Many iris recognition methods follow a quite common scheme, close to the well known and commercially most successful approach by J. Daugman: After image acquisition, in a first step the iris texture is localised and extracted. From this texture, discriminative features are derived, which then can be used for comparison. Since only the iris part of captured eye images is used for later feature extraction, it makes sense to consider only the iris textures in our image transformation experiments. We therefore always extract an iris texture from eye images as a first step. We assume the texture to be the area between the two almost concentric circles of the pupil and the outer iris. These two circles are found by contrast adjustment, followed by Canny edge detection and Hough transformation. After the circles are detected, unwrapping along polar coordinates is done to obtain a rectangular texture of the iris. Finally, we resample the data to a size of 512x64 pixels. Andreas Uhl 4
Visual Example: Iris Texture Andreas Uhl 5
Iris Recognition: Approach by Ma et al. The texture is divided into N stripes to obtain N one-dimensional signals, each one averaged from the pixels of M adjacent rows. We used N = 10 and M = 5 for our 512x64 pixel textures (only the 50 rows close to the pupil are used from the 64 rows). A dyadic wavelet transform is then performed on each of the resulting 10 signals, and two fixed subbands are selected from each transform. This leads to a total of 20 subbands. In each subband we then locate all local minima and maxima above some threshold, and write a bitcode alternating between 0 and 1 at each extreme point. Using 512 bits per signal, the final code is then 512x20 bit. Once bitcodes are obtained, matching can be performed on them using Hamming distance. For matching to work well, we compensate for eye tilt by shifting the bit-masks during matching and use the concept of a “noise mask” to take care of hidden or distorted parts of the iris. Andreas Uhl 6
Cancelable Iris Templates Using a key to seed a pseudo-random-number generator, we construct transformed versions of the iris textures. A significant advantage of this approach is, that conventional feature extraction and matching can be applied to the distorted textures. The transforms originally proposed by Ratha et al. are used. 1. Block Re-mapping : each block of the target texture is mapped to a block from the source texture. A re-mapping of blocks instead of a permutation should be preferred for the application of cancelable biometrics, as it is not reversible. Source blocks which are not part of the mapping are not contained in the transformed texture at all, and therefore it is impossible to reconstruct the complete original. 2. Image Warping : the texture is re-mapped according to a distorted grid mesh laid over it. A key is used to specify one particular distortion, by offsetting each vertex in the original mesh by some amount. This is done by starting with a regular grid placed over the texture, in which the vertices are then randomly displaced using the key as seed to a pseudo random number generator. Key-space sizes are discussed in the paper. Andreas Uhl 7
Visual Example: Cancelable Iris Templates Andreas Uhl 8
Experimental Settings Test Images: Interval dataset out of the CASIA Iris V3 database , consisting of 2653 images in 396 classes (i.e. persons). Software: custom C-implementation of algorithm by Ma et al. First Test: Matching Performance: we assigned a random key to each class, then calculated the Hamming distance of resulting bit-codes between any two images ( 3517878 iris comparisons, 9008 of which are intra-class comparisons). Second Test: Key Sensitivity: an iris class is copied multiple times, and each such class is then assigned a random key as before. If the key-dependent transformations don’t lead to sufficiently distinct features, in this case it will shows up as high FMR because features from different classes will match. We used the first 20 classes with at least 10 samples out of the Interval dataset, and created 50 random keys for each to have a roughly similar number of comparisons to the first test ( 2495000 iris comparisons, 45000 of which are intra-class). Measurements: ROC curves (FNMR vs. FMR) and EER. Andreas Uhl 9
Results: Block Re-mapping with squared Blocks 50 50 ROC (remapped, 1x1) ROC (remapped, keys, 1x1) ROC (remapped, 2x2) ROC (remapped, keys, 2x2) ROC (remapped, 4x4) ROC (remapped, keys, 4x4) ROC (remapped, 8x8) ROC (remapped, keys, 8x8) ROC (remapped, 16x16) ROC (remapped, keys, 16x16) ROC (remapped, 32x32) ROC (remapped, keys, 32x32) 40 40 ROC (untransformed) EER (8.669275) EER (7.285259) EER (5.938485) EER (5.559553) EER (17.162864) EER (17.601281) EER (7.796788) EER (10.349781) EER (5.086874) EER (7.022821) EER (0.847283) 30 30 EER (1.608968) EER (1.068704) FNMR (%) FNMR (%) 20 20 10 10 0 0 0.001 0.01 0.1 1 10 100 0.001 0.01 0.1 1 10 100 FMR (%) FMR (%) Matching Performance Key Sensitivity → Test 1 : original algorithm delivers an EER of about 1.1%. Using random − re-mappings of 32x32 pixel blocks (only 32 such blocks fit into the used 512x64 pixel textures), matching performance decreases, with a resulting EER of 1.6%. → Test 2 : The big 32x32 pixel blocks obtain an EER of 0.8%, which in this − case is an indicator for how distinct keys are from each other. Andreas Uhl 10
Results: Block Re-mapping with rectangular Blocks Test 1: Matching performance size(pixel) 56x7 64x8 73x9 85x10 102x12 128x16 blocks 81 64 49 36 25 16 EER (%) 1.3 1.6 1.2 1.2 1.2 1.6 Test 2: Key sensitivity size(pixel) 56x7 64x8 73x9 85x10 102x12 128x16 blocks 81 64 49 36 25 16 EER (%) 0.2 0.4 0.2 0.2 0.3 1.0 → Matching performance can almost be maintained (EER 1.2% vs. 1.1%). − → EER for Test 2 can be kept rather low, resulting in reasonable key-sensitivity. − Andreas Uhl 11
Results: Warping when Changing the Offset Range Fixed grid with 128 vertices, consisting of 16x16 pixel sized blocks, varying ranges for the horizontal and vertical pixel offsets. 50 50 ROC (warped, 32x4_1_1) ROC (warped, keys, 32x4_1_1) ROC (warped, 32x4_2_2) ROC (warped, keys, 32x4_2_2) ROC (warped, 32x4_4_4) ROC (warped, keys, 32x4_4_4) ROC (warped, 32x4_8_8) ROC (warped, keys, 32x4_8_8) ROC (warped, 32x4_16_16) ROC (warped, keys, 32x4_16_16) ROC (untransformed) EER (47.703867) 40 40 EER (1.140107) EER (34.545985) EER (1.159404) EER (11.030851) EER (1.089066) EER (4.184531) EER (1.644508) EER (5.823112) EER (6.250133) EER (1.068704) 30 30 FNMR (%) FNMR (%) 20 20 10 10 0 0 0.001 0.01 0.1 1 10 100 0.001 0.01 0.1 1 10 100 FMR (%) FMR (%) Matching Performance Key Sensitivity → Test 1 : offset in the range of 8 pixels increases the EER to 1.6%. − → Test 2 : best result with same settings: EER of 4%. − Andreas Uhl 12
Recommend
More recommend
