SafetyPage

Research papers · Paper

Statistically Undetectable Backdoors in Deep Neural Networks

Andrej Bogdanov, Alon Rosen, Neekon Vafa · 10 Jul 2026

ReadSource at arxiv.org ↗Read hereCC BY 4.0

Outline · 41 pages

  1. Introductionp. 3
  2. Our Resultsp. 3
  3. Interpretationsp. 6
  4. Cryptographic Assumptions & The Johnson-Lindenstrauss Lemmap. 7
  5. Related Workp. 8
  6. Overview of Our Constructionp. 9
  7. Backdooring Gaussian Matricesp. 9
  8. Concentration in the Number of Solutionsp. 11
  9. Backdoors in Neural Networksp. 12
  10. Preliminariesp. 12
  11. Divergencesp. 13
  12. Number Balancing and Symmetric Binary Perceptronsp. 14
  13. Backdoors for Random Gaussian Projectionsp. 15
  14. Sampling the Backdoorp. 16
  15. Concentration in the Number of Solutionsp. 18
  16. Putting It All Togetherp. 22
  17. Tightnessp. 23
  18. Constructing Backdoors for Neural Networksp. 24
  19. Defining Backdoorsp. 24
  20. Neural Network Preliminariesp. 26
  21. Constructionp. 27
  22. Backdoors in Deep Neural Networksp. 30
  23. Basic Implementation and Experimentsp. 32
  24. Proof of Concept Implementationp. 32
  25. Computational Hardness of Collision Findingp. 33
  26. Concluding Remarksp. 34