Concept-based attacks target neural-network logic, not class labels
Original title: Attack logics, not outputs: Concept-based Property Attacks for Shortcut Elimination in Neural Networks
Retraining on the attacks kept classical robustness gains while better removing semantic shortcuts.
Researchers argue that neural-network robustness analysis should falsify concept-level logical properties rather than only final class outputs. They formalize constraints over human-interpretable concepts and implement them on already trained networks using concept-based explainability methods. They prove these attacks have a smaller expected search space than class falsification. An empirical study across vision models and a code-processing task found retraining preserved robustness gains from classical adversarial training while better eliminating semantic biases.
- Classical attacks only falsify a final output class.
- Properties are applied to trained networks with C-XAI techniques.
- The study includes vision datasets and a code-processing task.
- Authors say the attacks align better with intuitive robustness targets.
From the original article
1 min read at arxiv.orgDeep neural networks (NNs) for computer vision are vulnerable to adversarial attacks, such that minuscule malicious changes to inputs may induce unintuitive outputs. One key approach to verify and mitigate such robustness issues is to falsify expected output behavior, which enables local robustness guarantees and adversarial training.
By Raik Dankworth · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.