Secure Self-supervised Learning: Challenges and Solutions

Secure Self-supervised Learning: Challenges and Solutions

Google TechTalks via YouTube Direct link

Summary

20 of 32

20 of 32

Summary

Class Central Classrooms beta

YouTube videos curated by Class Central.

Classroom Contents

Secure Self-supervised Learning: Challenges and Solutions

Automatically move to the next video in the Classroom when playback concludes

  1. 1 Intro
  2. 2 Conventional Paradigm: Supervised Learning
  3. 3 Key Challenge of Supervised Learning
  4. 4 Road Map
  5. 5 Background on Self-supervised Learning
  6. 6 Data Augmentation
  7. 7 Pre-training an Encoder - SimCLR [ICML'20]
  8. 8 Building a Downstream Classifier
  9. 9 Backdoor Attack
  10. 10 Key Idea of Our Attack
  11. 11 Quantifying Effectiveness Goal
  12. 12 Quantifying Condition I
  13. 13 Quantifying Utility Goal
  14. 14 Optimization Problem
  15. 15 Attack Setting
  16. 16 Attack Success Rate
  17. 17 Clean Accuracy and Backdoored Accuracy
  18. 18 Evaluation on Real-world Pre-trained Encoders
  19. 19 Existing Defenses are Insufficient
  20. 20 Summary
  21. 21 Motivation on Data Auditing
  22. 22 Auditing Unauthorized Data Use
  23. 23 Examples of Real-world Unauthorized Data Use
  24. 24 Our EncoderMI: Membership Inference based Data Auditing for Pre-trained Encoders
  25. 25 Revisiting Encoder Pre-training
  26. 26 Shadow Training Setup
  27. 27 Pre-training a Shadow Encoder
  28. 28 Constructing a Training Set for Inference Classifier
  29. 29 Building an Inference Classifier
  30. 30 Experimental Setup
  31. 31 Evaluation on CLIP
  32. 32 Conclusion

Never Stop Learning.

Get personalized course recommendations, track subjects and courses with reminders, and more.

Someone learning on their laptop while sitting on the floor.