This package implements a motion-based counter-measure to spoofing attacks to face recognition systems as described at the paper Counter-Measures to Photo Attacks in Face Recognition: a public database and a baseline, by Anjos and Marcel, International Joint Conference on Biometrics, 2011.
If you use this package and/or its results, please cite the following publications:
The original paper with the counter-measure explained in details:
@inproceedings{Anjos_IJCB_2011, author = {Anjos, Andr{\'{e}} and Marcel, S{\'{e}}bastien}, keywords = {Attack, Counter-Measures, Counter-Spoofing, Disguise, Dishonest Acts, Face Recognition, Face Verification, Forgery, Liveness Detection, Replay, Spoofing, Trick}, month = oct, title = {Counter-Measures to Photo Attacks in Face Recognition: a public database and a baseline}, booktitle = {International Joint Conference on Biometrics 2011}, year = {2011}, url = {http://publications.idiap.ch/downloads/papers/2011/Anjos_IJCB_2011.pdf} }
Bob as the core framework used to run the experiments:
@inproceedings{Anjos_ACMMM_2012, author = {A. Anjos and L. El Shafey and R. Wallace and M. G\"unther and C. McCool and S. Marcel}, title = {Bob: a free signal processing and machine learning toolbox for researchers}, year = {2012}, month = oct, booktitle = {20th ACM Conference on Multimedia Systems (ACMMM), Nara, Japan}, publisher = {ACM Press}, url = {http://publications.idiap.ch/downloads/papers/2012/Anjos_Bob_ACMMM12.pdf}, }
If you decide to use the REPLAY-ATTACK database, you should also mention the following paper, where it is introduced:
@inproceedings{Chingovska_BIOSIG_2012, author = {Chingovska, Ivana and Anjos, Andr{\'{e}} and Marcel, S{\'{e}}bastien}, keywords = {Attack, Counter-Measures, Counter-Spoofing, Face Recognition, Liveness Detection, Replay, Spoofing}, month = sep, title = {On the Effectiveness of Local Binary Patterns in Face Anti-spoofing}, booktitle = {IEEE Biometrics Special Interest Group}, year = {2012}, url = {http://publications.idiap.ch/downloads/papers/2012/Chingovska_IEEEBIOSIG2012_2012.pdf}, }
If you wish to report problems or improvements concerning this code, please contact the authors of the above mentioned papers.
This method was originally conceived to work with the the PRINT-ATTACK database, but has since evolved to work with the whole of the the REPLAY-ATTACK database, which is a super-set of the PRINT-ATTACK database. You are allowed to select protocols in each of the applications described in this manual. To generate the results for the paper, just select print as protocol option where necessary. Detailed comments about specific results or tables are given where required.
The data used in the paper is publicly available and should be downloaded and installed prior to try using the programs described in this package. The root directory of the database installation is used by the first program in the Antispoofing-Motion toolchain.
Note
If you are reading this page through our GitHub portal and not through PyPI, note the development tip of the package may not be stable or become unstable in a matter of moments.
Go to http://pypi.python.org/pypi/antispoofing.motion to download the latest stable version of this package.
There are 2 options you can follow to get this package installed and operational on your computer: you can use automatic installers like pip (or easy_install) or manually download, unpack and use zc.buildout to create a virtual work environment just for this package.
Using pip
is the easiest (shell commands are marked with a $
signal):
$ pip install antispoofing.motion
You can also do the same with easy_install
:
$ easy_install antispoofing.motion
This will download and install this package plus any other required dependencies. It will also verify if the version of Bob you have installed is compatible.
This scheme works well with virtual environments by virtualenv or if you have root access to your machine. Otherwise, we recommend you use the next option.
Download the latest version of this package from PyPI and unpack it in your working area. The installation of the toolkit itself uses buildout. You don't need to understand its inner workings to use this package. Here is a recipe to get you started:
$ python bootstrap.py $ ./bin/buildout
These 2 commands should download and install all non-installed dependencies and get you a fully operational test and development environment.
Note
The python shell used in the first line of the previous command set
determines the python interpreter that will be used for all scripts developed
inside this package. Because this package makes use of Bob, you must make sure that the bootstrap.py
script is called with the same interpreter used to build Bob, or
unexpected problems might occur.
If Bob is installed by the administrator of your system, it is safe to
consider it uses the default python interpreter. In this case, the above 3
command lines should work as expected. If you have Bob installed somewhere
else on a private directory, edit the file buildout.cfg
before
running ./bin/buildout
. Find the section named buildout
and edit or
add the line prefixes
to point to the directory where Bob is installed or
built. For example:
[buildout] ... prefixes=/Users/crazyfox/work/bob/build
It is assumed you have followed the installation instructions for the package
and got this package installed and the REPLAY-ATTACK (or PRINT-ATTACK) database
downloaded and uncompressed in a directory to which you have read access.
Through this manual, we will call this directory /root/of/database
. That
would be the directory that contains the sub-directories train
, test
,
devel
and face-locations
.
At Idiap, we use the powerful Sun Grid Engine (SGE) to parallelize our job
submissions as much as we can. At the Biometrics group, we have developed a
little toolbox <http://pypi.python.org/pypi/gridtk> that can submit and
manage jobs at the Idiap computing grid through SGE. If you are at Idiap, you
can download and install this toolset by adding gridtk
at the eggs
section of your buildout.cfg
file, if it is not already there. If you are
not, you still may look inside for tips on automated parallelization of
scripts.
The following sections will explain how to reproduce the paper results in
single (non-gridified) jobs. A note will be given where relevant explaining how
to parallalize the job submission using gridtk
.
The first stage of the process is to calculate the normalized frame differences
using video sequences. The program that will do that should be sitting in
bin/motion_framediff.py
. It can calculate normalize frame differences in distinct
parts of the scene (given you provide face locations for each of the frames in
all video sequences to be analyzed).
To execute the frame difference process to all videos in the REPLAY-ATTACK database, just execute:
$ ./bin/motion_framediff.py /root/of/database results/framediff replay
There are more options for the motion_framediff.py
script you can use (such
as the sub-protocol selection for the Replay Attack database). Note that, by
default, all applications are tunned to work with the whole of the
database. Just type --help
after the keyword replay
at the command
line for instructions.
Note
To parallelize this job, do the following:
$ ./bin/jman submit --array=1200 ./bin/motion_framediff.py /root/of/database results/framediff replay
The magic number of 1200 entries can be found by executing:
$ ./bin/motion_framediff.py --grid-count replay
Which just prints the number of jobs it requires for the grid execution.
The second step in calculating the frame differences is to compute the set of 5 quantities that are required for the detection process. To reproduce the results in the paper, we accumulate the results in windows of 20 frames, without overlap:
$ ./bin/motion_diffcluster.py results/framediff results/quantities replay
There are more options for the motion_diffcluster.py script you can use (such as the sub-protocol selection). Just type --help at the command line for instructions.
Note
This job is very fast and normally does not require parallelization. You can still do it with:
$ ./bin/jman submit --array=1200 ./bin/motion_diffcluster.py results/framediff results/quantities replay
Training a linear machine to perform LDA should go like this:
$ ./bin/motion_ldatrain.py --verbose results/quantities results/lda replay
This will create a new linear machine train it using the training data. Evaluation based on the EER on the development set will be performed by the end of the training:
Performance evaluation: -> EER @ devel set threshold: 8.11125e-02 -> Devel set results: * FAR : 16.204% (175/1080) * FRR : 16.174% (558/3450) * HTER: 16.189% -> Test set results: * FAR: 16.389% (236/1440) * FRR: 18.641% (856/4592) * HTER: 17.515%
The resulting linear machine will be saved in the output directory called
results/lda
.
Training MLPs to perform discrimination should go like this:
$ ./bin/motion_rproptrain.py --verbose --epoch=10000 --batch-size=500 --no-improvements=1000000 --maximum-iterations=10000000 results/quantities results/mlp replay
This will create a new MLP and train it using the data produced by the "clustering" step. The training can take anywhere from 20 to 30 minutes (or even more), depending on your machine speed. You should see some debugging output with the partial results as the training go along:
... iteration: RMSE:real/RMSE:attack (EER:%) ( train | devel ) 0: 9.1601e-01/1.0962e+00 (60.34%) | 9.1466e-01/1.0972e+00 (58.71%) 0: Saving best network so far with average devel. RMSE = 1.0059e+00 0: New valley stop threshold set to 1.2574e+00 10000: 5.6706e-01/4.2730e-01 (8.29%) | 7.6343e-01/4.3836e-01 (11.90%) 10000: Saving best network so far with average devel. RMSE = 6.0089e-01 10000: New valley stop threshold set to 7.5112e-01 20000: 5.6752e-01/4.2222e-01 (8.21%) | 7.6444e-01/4.3515e-01 (12.07%) 20000: Saving best network so far with average devel. RMSE = 5.9979e-01 20000: New valley stop threshold set to 7.4974e-01
The resulting MLP will be saved in the output directory called
results/mlp
. The resulting directory will also contain performance
analysis plots. The results derived after this step are equivalent to the
results shown at Table 2 and Figure 3 at the paper.
To get results for specific supports as shown at the first two lines of Table
2, just select the support using the --support=hand
or --support=fixed
as a flag to motion_rproptrain.py
. Place this flags after the keyword
replay
at the command line. At this point, it is adviseable to use
different output directories using the --output-dir
flag as well. If you
need to modify or regenerate Figure 3 at the paper, just look at
antispoofing/motion/ml/perf.py
, which contains all plotting and analysis
routines.
Note
If you think that the training is taking too long, you can interrupt it by
pressing CTRL-C
. This will cause the script to quit gracefully and still
evaluate the best MLP network performance to that point.
Note
To execute this script in the grid environment, just set the output directory to depend on the SGE_TASK_ID environment variable:
$ ./bin/jman --array=10 ./bin/motion_rproptrain.py --verbose --epoch=10000 --batch-size=500 --no-improvements=1000000 --maximum-iterations=10000000 results/quantities 'results/mlp.%(SGE_TASK_ID)s' replay
You should now dump the scores for every input file in the
results/quantities
directory using the motion_make_scores.py
script,
for example, to dump scores produced with by an MLP:
$ ./bin/motion_make_scores.py --verbose results/quantities results/mlp/mlp.hdf5 results/mlp-scores replay
This should give you the detailed output of the machine for every input file in the training, development and test sets. You can use these score files in your own score analysis routines, for example.
Note
The score file format is an HDF5 file with a single array, which contains the scores for every frame in the input video. Values which are marked as NaN should be ignored by your procedure. The reason varies: it may mean no valid face was detected on such a frame or that the motion-detection procedure decided to skip (on user configuration) the analysis of that frame.
The time analysis is the end of the processing chain, it fuses the scores of instantaneous outputs to give out a better estimation of attacks and real-accesses for a set of frames. You can used with the scores output by MLPs or linear machines (LDA training). To use it, write something like:
$ ./bin/motion_time_analysis.py --verbose results/mlp-scores results/mlp-time replay
The 3 curves on Figure 4 at the paper relate to the different support types.
Just repeat the procedure for every system trained with data for a particular
support (equivalent for then entries in Table 2). To set the support use
--help
after the keyword replay
on the command-line above to find out
how to specify the support to this program. The output for this script is
dumped in PDF (plot) and text (.rst
file) on the specified directory.
If you wish to create a single 5-column format file
by combining this counter-measure scores for every video into a single file
that can be fed to external analysis utilities such as our
antispoofing.evaluation <http://pypi.python.org/pypi/antispoofing.evaluation>
package, you should use the script motion_merge_scores.py
. You will have to
specify how many of the scores in every video you will want to average and the
input directory containing the scores files that will be merged.
The output of the program consists of three 5-column formatted files with the client identities and scores for every video in the input directory. A line in the output file corresponds to a video from the database.
You run this program on the output of motion_make_scores.py
. So, it should
look like this if you followed the previous example:
$ ./bin/motion_merge_scores.py results/mlp-scores results/mlp-merged replay
The above commandline examples will generate 3 files containing the training, development and test scores, accumulated over each video in the respective subsets, for input scores in the given input directory.
In case of problems, please contact any of the authors of the paper.