Reproduce “Transferability in Machine Learning: from Phenomena to Black-Box Attacks using Adversarial Samples” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nicolas Papernot, Patrick McDaniel, Ian Goodfellow
This paper establishes that adversarial examples crafted for one machine learning model can successfully fool other models, a property termed 'transferability', thereby enabling practical black-box attacks. The research quantifies this phenomenon across multiple machine learning techniques and demonstrates its application by achieving high misclassification rates against commercial machine learning platforms.
Key insights
- Adversarial examples transfer effectively between machine learning models, including those with different architectures or training data, a property termed 'transferability'.
- All tested machine learning techniques (DNNs, Logistic Regression, SVMs, Decision Trees, kNN) exhibit vulnerability to adversarial examples, varying in degree.
- Practical black-box attacks against deployed machine learning systems are feasible, requiring only query access to the target model and demonstrating threat to commercial platforms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.