Team for Research in
Ubiquitous Secure Technology

Citation
Alvaro Cardenas. "Evaluation of Classifiers used in Security Applications". Talk or presentation, 24, May, 2006.

Abstract
In recent years several tools based on statistical methods and machine learning have been incorporated in security related tasks involving classification, such as intrusion detection systems, fraud detection, spam filters, biometrics and multimedia forensics. Measuring the security performance of these classifiers is an essential part for facilitating decision making, determining the viability of the product, and providing the insight necessary to improve the design of the system. There are however relevant considerations for security related problems that are sometimes ignored by traditional evaluation schemes. The first consideration is the usually large class imbalance between normal events and attack events. The second consideration is the fact that the classifier or learning rule will be deployed in an adversarial environment. In this talk we introduce the Bayesian-ROC curves for the class imbalance problem and provide a framework to evaluate the performance under the worst type of adversarial attacks. We provide practical examples in intrusion detection and multimedia watermarking.

Electronic downloads

• Cardenas_EvalOfClassifiersUsedInSecurityApps_TrustSeminar_05_24_06.pdf · application/pdf · 3872 kbytes

• HTML

   Alvaro Cardenas. <a
  href="http://www.truststc.org/pubs/102.html"
  ><i>Evaluation of Classifiers used in Security
  Applications</i></a>, Talk or presentation,  24,
  May, 2006.

• Plain text

   Alvaro Cardenas. "Evaluation of Classifiers used in
  Security Applications". Talk or presentation,  24, May,
  2006.

• BibTeX

  @presentation{Cardenas06_EvaluationOfClassifiersUsedInSecurityApplications,
      author = { Alvaro Cardenas},
      title = {Evaluation of Classifiers used in Security
                Applications},
      day = {24},
      month = {May},
      year = {2006},
      abstract = {In recent years several tools based on statistical
                methods and machine learning have been
                incorporated in security related tasks involving
                classification, such as intrusion detection
                systems, fraud detection, spam filters, biometrics
                and multimedia forensics. Measuring the security
                performance of these classifiers is an essential
                part for facilitating decision making, determining
                the viability of the product, and providing the
                insight necessary to improve the design of the
                system. There are however relevant considerations
                for security related problems that are sometimes
                ignored by traditional evaluation schemes. The
                first consideration is the usually large class
                imbalance between normal events and attack events.
                The second consideration is the fact that the
                classifier or learning rule will be deployed in an
                adversarial environment. In this talk we introduce
                the Bayesian-ROC curves for the class imbalance
                problem and provide a framework to evaluate the
                performance under the worst type of adversarial
                attacks. We provide practical examples in
                intrusion detection and multimedia watermarking.},
      URL = {http://www.truststc.org/pubs/102.html}
  }
  

Posted by Christopher Brooks on 2 Jun 2006.
Groups: trustseminar
For additional information, see the Publications FAQ or contact webmaster at www truststc org.