Security of Quantum Machine Learning Systems
summary
Claims that quantum models need only ~10^2–10^3 labelled images for comparable medical classification, achieved by swapping a classical conv layer for a quantum convolutional layer, are supported by theoretical arguments about expressivity and manifold curvature and by prototype hybrid circuits using QBLAS, HHL and quantum kernel subroutines. New attack vectors target the encoding step and device noise: attackers can insert gates to break the encoding or mount a crosstalk attack across co‑located circuits, while studies show random noise can sometimes increase robustness and adversarial examples transfer between classical and quantum models.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:35Presentation
- 6:36Example
- 9:40Classical AI Security
- 11:51Quantum Machine Learning Security
- 14:41Encoding
- 15:23Encoding examples
- 16:08Crosstalk
- 17:02Ion Trap
- 18:19Conclusion
- 19:01Questions
related talks
44mQuantum machine learning: Prospects and challenges
26mBeyond the Hype: Hybrid Quantum-Classical AI Agents for Scientific Discovery
29mA Rigorous and Robust Quantum Speed-up in Supervised Machine Learning
19mUsing hybrid quantum-classical convolutional neural networks to id
29m10 Things You Should Know About Quantum Computing and ML
01h 06mPractical Quantum Machine Learning: Overcoming the Bottlenecks of Scalability and(…)
34mDisease Detection in the Age of Quantum Computing
46mQuantum Machine Learning 101: An Introduction for Beginners with...