Who is behind these predictions Reconciling transparency and privacy in machine learning
Abstract
In this project we will analyse whether it is possible to capture the (near-)exact machine learning family used for training the model attacked (e.g., decision stumps, decision trees, set of rules or linear discriminants) as well as other intrinsic characteristics (feature space significance, transformations between features, context or behavioural changes, etc.) from a given relevant subset of queries. We will present appropriate query generation strategies that analyse both the relevance of the attributes with some levels of error and confidence, as well as the most valuable attributes according to the attacker’s value function. Finally, we will explore several countermeasures to reduce the effectiveness of the attacks, including attribute transformations (univariate and multivariate) or context or behavioural changes (such as cost matrices adjustments, threshold modifications, etc.) Furthermore, we will define metrics of internal interpretability and of external obfuscation (and the optimal trade-off between them).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 28, 2017
- Source ID
- FA95501710287
Entities
People
- Maria Jose Ramirez Quintana
Organizations
- Air Force Office of Scientific Research
- Technical University of Valencia
- United States Air Force