Active Learning Over a Network-Based DDDAS for Anomaly Detection and Class Discovery
Abstract
Main Accomplishments/Successes: A. Novel powerful techniques were tested for confidently detecting or certifying against imperceptible backdoor data poisoning of deep neural network image classifiers. B. Novel powerful techniques were developed for confidently detecting or certifying perceptible but scene-plausible backdoor data poisoning of neural network image classifiers. C. Novel techniques were developed for confidently detecting or certifying against data poisoning attacks on training datasets and sanitizing the data, to remove the data-poisoned training samples. D. Novel techniques were developed for detecting reverse engineering attacks on classifiers.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 25, 2021
- Accession Number
- AD1153896
Entities
People
- David A. B. Miller
Organizations
- Pennsylvania State University