Neural Network Anomaly Detection
Abstract
Deep neural networks (DNNs) are used across various domains and have the potential to reduce operational costs, increase productivity, and improve safety and security. Although DNNs are used with a high rate of success, they are inherently vulnerable to adversarial examples (AEs), which are inputs to a trained DNN that are modified so that the DNN is deceived into misclassifying the inputs. If this vulnerability is exploited, the DNN can produce unexpected results that may adversely affect public health and safety. In this research, analytical approaches identified image-based AEs. DNN models were created using the Tensorflow library. The hidden layer activations of "normal" and adversarial inputs were created using images from the Canadian Institute for Advanced Research-10 classes dataset. The Uniform Manifold Approximation and Projection code was used to visualize how AEs differed from normal inputs relative to class manifolds. The Stochastic Gradient Descent One Class Support Vector Machine anomaly detection algorithm was applied to the hidden layer activations to detect AEs. Resulting key insights into how DNNs interpreted AEs will lead to mathematical strategies to identify AEs.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 31, 2023
- Accession Number
- AD1221033
Entities
People
- Sherida T. Jacob
Organizations
- Naval Undersea Warfare Center