1. Self Adaptive Systems: An Experimental
Analysis of the Performance Over Time
Ajita Rattani, Gian Luca Marcialis, Fabio Roli
2. Biometric Verification Systems
They operate in two distinct stages:
Image is acquired for each user
(gallery) in controlled environment
1) The enrolment stage (for instance ISO/ IEC FCD 19794-5 standard)
2) The verification stage. âTemplateâ is created and identity
labels assigned
ďThe performance of biometric systems degrades quickly when input images
exhibit substantial variations compared to the enrolled âtemplatesâ
ďś Some face examples showing intra-class variations in input data un-
represented by enrolled template
TEMPLATE QUERY IMAGES
2
3. Initial Attempts to Increase Template
Representativeness
⢠Re-enrollment
⢠Multibiometrics (Handbook of Multi-biometrics, 2006)
⢠Virtual biometric template synthesis (pose correction,
illumination correction, de-ageing transformations) (Wang et
al. 2006, Geng et al. 2007)
3
4. Recent Introduction-
Template update Methods
⢠Characteristics:
â Adapt themselves to the intra-class variation of the input
data.
â Minimizing performance loss due to unrepresentative
and outdated templates
⢠Commonly adopted is self-adaptive systems.
4
5. Self-Adaptive Systems
⢠Highly confidently
classified samples are
used for adaptation
⢠In order to avoid
impostor introduction
⢠Claimed to be robust
against short and
medium term intra-
class variations
6.
7. State-of-the-art
Reference Modality Impostor Database
X. Jiang Finger No 100x8
and W. Ser 12x200
Roli et al Face No 100x8
Ryu et al. Finger No 41x100
Pavani et Face No 5 months
al.
Till date: No paper has shown the performance robustness
over time
Reason: Unavailability of large number of samples collected
over a period of time, per user basis
Assumption of absence of impostor
No theoretical explanation of the functioning
8. Contributions
â This is the first study evaluating the
performance of self-adaptive systems, on the
input batch of samples as available over time
â The conceptual explanation of the functioning
of self-adaptive systems, supported by
experimental validations
â DIEE multimodal database has been explicitly
collected for this aim, over a span of 1.5 years
9. Conceptual representation
A hypothetical diagram showing the The representational capability of each
initial condition where the enrolled template in the updated set on
template is shown with the help of star adaptation using samples 1, 2 and 3
and encircled in its representation region
10. Contd...
As a result overall genuine region In the real time environment impostor
expands samples may also be present
11. Experimental Validations
⢠Dataset: DIEE Multimodal database
â 49 subjects with 50 samples per subject acquired in five
sessions with 10 samples per session
â Acquired in a time span of 1.5 years
â Containing temporal as well as other intra-class variations
Example facial images taken from two different
sessions for a randomnly choosen user
12. Experimental Protocol
⢠Training: 2 enrolled images per person from
the batch b_1
⢠Updating: batches two to four
⢠Performance evaluation:On updating
using batch b_i performance is evaluated
for batch b_i+1and EER_i computed
13. Experiment #1
⢠Aim: to evaluate the performance of self adaptive
systems over time
⢠Assumption of absence of impostorâs access.
⢠Updating threshold: 0.01 % FAR
14. Results
Performance
enhancement
and stability can
be attained over
time
Performance of self-adaptive systems for index and
thumbprint biometrics in comparison to baseline classifier
15. At varying threshold conditions
Large variation in
the performance
from one
updating cycle to
another as a
result of
representation
region expansion
significantly
Performance of self-adaptive face recognition system
at varying thresholds from stringent to relaxed for face
biometrics
16. Table: showing percentage of samples gradually
added to the userâs gallery for face biometrics
Threshold (%) Cycle 1 (%) Cycle 2 (%) Cycle 3 (%) Cycle 4
at % FAR
0.00001 % 31.17 21.60 19.75 19.36
0.00001 % 33.3 26.54 26.95 27.70
0.01 % 52.16 55.86 61.83 65.74
0.1 % 56.79 62.03 68.31 70.98
1% 60.8 68.51 73.66 75.30
17. Further confirmation:- representation
region expansion
The scores obtained on fifth batch using the The scores obtained on fifth batch using the
baseline matcher enrolled with two templates self-adaptive system updated using 1 to 4
from a random user for thumb biometric batches for a random user for thumb biometric
18. Experiment #2
⢠Aim: is to evaluate the performance of self adaptive systems over
time on the assumption of presence of impostorâs access.
⢠Assumption: Presence of impostors, enabling the evaluation of
impostorâs intrusion over time
⢠Updating Threshold: 0.001 % FAR
19. â˘Performance
strongly suffers
from the operation
at stringent
threshold
â˘However, the
stability can be
obtained over time
â˘Different
biometrics show
difference trend in
performance
improvement
Performance of self adaptive thumb and index
fingerprint system under the assumption of
impostor presence
20. Conclusions and Future work
⢠Self-adaptive systems can result in performance enhancement and
classifierâs stability over time
⢠The obtained performance enhancement is very much dependent
on the set updating threshold
⢠Different biometric may show different performance trend over time.
⢠The possibility of presence and updating due to impostor is a
serious and open issue.
⢠Future work will rely on further development of the conceptual
behaviour and significant in-depth analysis of impostorâs effect over
time.