Team for Research in
Ubiquitous Secure Technology

Citation
Marco Barreno, Blaine Nelson, Russell Sears, Anthony D. Joseph, J. D. Tygar. "Can Machine Learning Be Secure?". Talk or presentation, 27, April, 2006; Poster Given at the NSF Trust Site Visit.

Abstract
Statistical learning is an invaluable tool that is increasingly being used in security-sensitive applications, but little attention has been paid to the possibility that new vulnerabilities may be introduced by learning systems. We investigate a broad class of statistical learning algorithms and show that their use creates new potential vulnerabilities that an attacker may be able to exploit. We discuss and analyze the range of potential attacks and their effects. We also explore defenses along the lines of adding robustness to the algorithms and selecting appropriate algorithms and model parameters in the first place. Finally, we present some theoretical analysis and experiments to evaluate the attacks and defenses.

Electronic downloads

• barreno_MachineLearning_TrustApril06Poster.pdf · application/pdf · 315 kbytes

• HTML

   Marco Barreno, Blaine Nelson, Russell Sears, Anthony D.
  Joseph, J. D. Tygar. <a
  href="http://www.truststc.org/pubs/71.html"
  ><i>Can Machine Learning Be
  Secure?</i></a>, Talk or presentation,  27,
  April, 2006; Poster Given at the NSF Trust Site Visit.

• Plain text

   Marco Barreno, Blaine Nelson, Russell Sears, Anthony D.
  Joseph, J. D. Tygar. "Can Machine Learning Be
  Secure?". Talk or presentation,  27, April, 2006;
  Poster Given at the NSF Trust Site Visit.

• BibTeX

  @presentation{BarrenoNelsonSearsJosephTygar06_CanMachineLearningBeSecure,
      author = { Marco Barreno, Blaine Nelson, Russell Sears,
                Anthony D. Joseph, J. D. Tygar},
      title = {Can Machine Learning Be Secure?},
      day = {27},
      month = {April},
      year = {2006},
      note = {Poster Given at the NSF Trust Site Visit},
      abstract = {Statistical learning is an invaluable tool that is
                increasingly being used in security-sensitive
                applications, but little attention has been paid
                to the possibility that new vulnerabilities may be
                introduced by learning systems. We investigate a
                broad class of statistical learning algorithms and
                show that their use creates new potential
                vulnerabilities that an attacker may be able to
                exploit. We discuss and analyze the range of
                potential attacks and their effects. We also
                explore defenses along the lines of adding
                robustness to the algorithms and selecting
                appropriate algorithms and model parameters in the
                first place. Finally, we present some theoretical
                analysis and experiments to evaluate the attacks
                and defenses.},
      URL = {http://www.truststc.org/pubs/71.html}
  }
  

Posted by Christopher Brooks on 4 May 2006.
Groups: trust
For additional information, see the Publications FAQ or contact webmaster at www truststc org.