Overview
Explore a comprehensive video explanation of the research paper "Radioactive data: tracing through training" in this 36-minute tutorial. Dive into the concept of marking datasets with hidden "radioactive" tags to detect their usage in training neural classifiers. Learn about the mechanics of neural classifiers, radioactive marking techniques, high-dimensional random vectors, backpropagation of fake features, and feature space realignment. Examine experimental results, including black-box testing, and gain insights into the implications for data privacy and model training. Understand how this method offers improved signal-to-noise ratio compared to data poisoning and backdoor approaches, with the ability to detect radioactive data usage even when only 1% of the training data is marked.
Syllabus
- Intro & Overview
- How Neural Classifiers Work
- Radioactive Marking via Adding Features
- Random Vectors in High-Dimensional Spaces
- Backpropagation of the Fake Features
- Re-Aligning Feature Spaces
- Experimental Results
- Black-Box Test
- Conclusion & My Thoughts
Taught by
Yannic Kilcher