Robust and Reversible Deep Nets for Synthetic Aperture Sonar Analysis
Abstract
Deep Learning approaches to Synthetic Aperture Sonar analysis have shown promise to serve as the basis for operational algorithms. However, it is well documented that Deep Networks are not robust to adversarial or outlier samples in test data and are not “competency aware”, i.e., they do produce a reliability score. The explainability of Deep Networks is limited which prevents understanding of network decision making . Mitigating these problems is required before using Deep Learning as the basis for a high-performance, trustworthy fielded system. To this end, we propose to investigate Null Space and Fourier Analysis methods for SAS imagery in network architectures. Every layer in a deep learning architecture has an associated null space which disregards information during analysis. Some information should be disregarded whereas other disregarding information can lead to false alarms or incorrect classifications. Current training techniques do not include the null space; it is not well understood how to characterize and control null space information processing. We will investigate partitions of the null space and the relationship between partitions, SAS image characteristics, and network performance. The investigation will be used to devise techniques that minimize null space projections and providing a measure of the likelihood a sample is adversarial for test samples, namely, the magnitude of the projection onto the null space. To enhance explainability of convolutional neural networks, invertible convolutions will be learned using Fourier transforms. The learning process requires constraining values of the frequency domain filter to be significantly larger than zero. It also requires mean pooling, and invertible activations. Inverting CNNs provides mechanisms for characterizing the set of inputs that can mapped to an output, yielding better performance characterization, and for identifying the salient components of input samples that drive network decisions and outputs.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 18, 2025
- Source ID
- N001742010015
Entities
People
- Alina Glenn
Organizations
- United States Navy
- University of Florida