Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Adversarial Examples: Let's fool modern AI systems with Physical Stickers.

Adversarial Examples: Let's fool modern AI systems with Physical Stickers.

A quick and dirty introduction to Adversarial Examples in Deep Learning.

Video: https://youtu.be/BpNET8N4dzQ

89c0f610f1ffe542cce1d7aef67e0212?s=128

Anant Jain

April 29, 2019
Tweet

Transcript

  1. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com Adversarial Examples:
 Let’s fool modern

    A.I. Systems! 1 Anant Jain Co-founder, CommonLounge.com (Compose Labs) https://commonlounge.com
 https://index.anantja.in @anant90
  2. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 2 What are Adversarial Examples?

  3. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 3

  4. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 4

  5. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 5 Three remarkable things about

    Adversarial examples
  6. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 6 Three remarkable things about

    Adversarial examples • Small perturbation • Amount of noise added is imperceivable
  7. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 7 Three remarkable things about

    Adversarial examples • Small perturbation • Amount of noise added is imperceivable • High Confidence • It was easy to attain high confidence in the incorrect classification
  8. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 8 Three remarkable things about

    Adversarial examples • Small perturbation • Amount of noise added is imperceivable • High Confidence • It was easy to attain high confidence in the incorrect classification • Transferability • Didn’t depend on the specific ConvNet used for the task.
  9. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 9 https://arxiv.org/pdf/1312.6199.pdf

  10. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 10 Take a correctly classified

    image (left image in both columns), and add a tiny distortion (middle) to fool the ConvNet with the resulting image (right).
  11. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 11 https://arxiv.org/pdf/1412.6572.pdf

  12. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 12

  13. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 13 https://arxiv.org/pdf/1607.02533.pdf

  14. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 14 Adversarial examples can be

    printed out on normal paper and photographed with a standard resolution smartphone and still cause a classifier to, in this case, label a “washer” as a “safe”.
  15. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 15 https://arxiv.org/pdf/1712.09665.pdf

  16. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com Demo 16

  17. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com Demo 17

  18. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com What are the implications of

    these attacks? 18
  19. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com What are the implications of

    these attacks? 19 •Self Driving Cars: A patch may make a car think that a Stop Sign is a Yield Sign.
  20. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com What are the implications of

    these attacks? 20 •Self Driving Cars: A patch may make a car think that a Stop Sign is a Yield Sign. •Alexa: Voice-based Personal Assistants: Transmit sounds that sound like noise, but give specific commands.
  21. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com What are the implications of

    these attacks? 21 •Self Driving Cars: A patch may make a car think that a Stop Sign is a Yield Sign. •Alexa: Voice-based Personal Assistants: Transmit sounds that sound like noise, but give specific commands. •Ebay: Sell livestock and other banned substances.
  22. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 22 125,478 views | Apr

    1, 2019, 07:06am Hackers Use Little Stickers To Trick Tesla Autopilot Into The Wrong Lane Cybersecurity I cover crime, privacy and security in digital and physical forms. Thomas Brewster Forbes Staff
  23. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com 23

  24. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com “Resistance to adversarial examples and

    figuring out how to make machine learning secure against an adversary who wants to interfere and control it is one of the most important problems researchers today could solve.” — Ian Goodfellow, inventor of GANs, April 2019. 24 What’s next?
  25. twitter.com/@anant90 Anant Jain, Co-Founder, CommonLounge.com Thanks for listening to this

    ⚡talk! 25 Anant Jain Co-founder, CommonLounge.com (Compose Labs) https://commonlounge.com/pathfinder
 https://index.anantja.in CommonLounge.com is an online-learning platform similar to Coursera/Udacity, except our courses are in the form of lists of text-based tutorials, quizzes and step- by-step projects instead of videos. 
 
 Check out our Deep Learning Course!