Advanced Seminar / Masterseminar SS17

The Advanced Seminar will focus on Advanced Topics in Biometrics.

Der Starttermin findet am 06.04.2017 um 10:15 im Raum D14/303 statt.

The seminar will have only few fixed class meetings. Besides these meetings, additional appointments shall be arranged individually and on demand. The schedule of this seminar can be found in the online booking system (OBS).

Each group’s term paper has to be prepared using Springer’s LNCS template with a length of 10 – 12 pages without references and clear marked appendices. The LNCS templates for Word and LaTeX can be found here: http://www.springer.com/computer/lncs?SGWID=0-164-6-793341-0. The final presentation will be 40 minutes per group + 20 minutes discussion of the results. A grade will be given based on the term paper and final presentation as defined in the module description.

According to OBS, the course capacity is 12 seats. Eventually, additional seats can be offered. If you don’t get a seat in the allocation period, please join the kickoff and ask for an additional seat. If you got allocated a seat but can’t join the kickoff, please get in touch with me by email before the kickoff. Otherwise, your seat will be revoked.

Student groups (1-2 persons) may select their own topic or choose a topic from the following list:

  1. Cross-spectral ocular biometrics: with the availability of different types of ocular sensors that are deployed for iris and periocular recognition under different illumination, significant degradation in biometric performance is expected when comparing such ocular images acquired under two different domains. Open-source biometric recognition software should be used on public available cross-spectral ocular database to identify the best performing (fusion of) features for cross-spectral recognition.
  2. Workload reduction in biometric identification systems: The quickly growing size of biometric deployments (e.g. the Indian Aadhaar project) confers many diverse challenges, one of which is system effciency in the identification mode. A naive implementation of a biometric system in identication mode requires 1:N template comparisons for a lookup. As the number of enrolled subjects (N) increases, the computational load and probability of false positive occurrences quickly become unacceptable. Based on a thorough literature study on workload reduction with respect to fingerprints or face students should compose a comprehensive survey summarizing, comparing and discussing results of existing works in the field. For this topic the term paper should be at least 16 pages.
  3. Unified multi-modality biometric feature space (proof-of-concept study): in multi-modal scenarios (e.g. face, signature, voice, iris), system fusions play an important role in order to benefit from information of each sub-system. By targeting feature space fusions, one method is to combine state-of-the-art features, however one may also engage a unified biometric feature extraction scheme based on modality-dependent signal processing front-ends. In a proof-of-concept study, implementations should be carried out utilizing state-of-the-art open source tools, targeting the intermediate-sized vector (i-vector) approach, such that an evaluation can be conducted about the pros/cons of different feature-space fusion strategies.
  4. Detecting voice imitation and disguises: presentation attacks pose a vital threat to biometric systems. Many studies put emphasis on machine-based attack schemes, whereas human based attacks are relevant, too. In order to detect mimicry and disguise, a database should be collected based on publicly available data (e.g. YouTube) from a broad range of suitable scenarios, such as politic satire and dub speaker featuring TV series, animes, movies, or games (with data of same dub speakers in different characters). A brief experimental study should be conducted based on promising open source tools for the purpose of providing insights on future research perspectives regarding human based presentation attacks on speaker recognition systems.
  5. Finger Presentation Attack Detection: the wide deployment of biometric recognition systems has also lead to the development of different attacking strategies to impersonate a subject. One of those strategies is known as Presentation Attack, and consists on presenting the system with an artefact which imitates the subject’s biometric characteristic (e.g., gummy fingers or video replays on a tablet). Based on a thorough literature study on Presentation Attack Detection (PAD) techniques (also known as anti-spoofing or liveness detection) for finger characteristics (e.g., fingerprint, finger vein), students should compose a comprehensive survey summarising, comparing and discussing results of existing works in the field. For this topic the term paper should be at least 16 pages.