Neural Network OOD Detection via Manifold Embedding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models, particularly deep convolutional neural networks (DCNNs), are vulnerable to out-of-distribution (OOD) samples, which can lead to adversarial attacks and compromise the trustworthiness and robustness of these models in real-world applications, as they often assign high confidence scores to OOD inputs, making it difficult to distinguish between in-distribution and OOD samples.
Innovation Solution
A method and system that embeds training data into a lower-dimensional space using manifold embedding techniques like Isometric Mapping (Isomap) and Locally Linear Embedding (LLE), and employs scoring mechanisms such as Expectation Maximization (EM) and Particle Swarm Optimization (PSO) to identify OOD samples and optimize the input space, thereby detecting attack surfaces and reducing the confidence scores of OOD samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are trained to recognize patterns, then their accuracy on in-distribution samples is improved, but their ability to distinguish out-of-distribution samples deteriorates as they assign high confidence scores to OOD inputs
Solution Approach 1:
The patent transforms the input data from the original high-dimensional space into a lower-dimensional embedded space using manifold embedding techniques (Isomap, LLE, or MDS). This dimensional transformation creates a new representation where OOD samples naturally separate from the training data distribution, enabling the model to distinguish OOD samples while maintaining accuracy on in-distribution samples. The embedded space provides an additional dimension for discrimination without sacrificing original classification performance.
2Ease of operation
If the model processes all input samples uniformly, then simplicity of operation is maintained, but detection precision of OOD samples deteriorates
Solution Approach 1:
The patent segments the input space into two distinct regions: in-distribution samples that fall within the embedded training data distribution, and out-of-distribution samples that fall outside this distribution. By computing a distance metric (Mahalanobis distance or Euclidean distance) from each sample to the embedded training distribution, the system creates a clear segmentation boundary. This allows uniform processing of all samples through the same embedding and distance computation pipeline while achieving precise OOD detection through the inherent segmentation of the embedded space.
3Productivity
If manifold embedding techniques are applied to reduce dimensionality, then computational efficiency is improved, but embedding precision may deteriorate due to information loss
Solution Approach 1:
The patent employs multiple different manifold embedding techniques (Isomap, Locally Linear Embedding, and Multidimensional Scaling) that use different parameterizations and optimization criteria. By providing alternatives with different computational complexities and preservation characteristics, the system can select the appropriate embedding method based on the specific requirements of the application. This parameter change approach allows balancing between computational efficiency and embedding precision by choosing the technique that best suits the data characteristics and performance needs.
Data Source
AI summary
There is provided systems and methods for counteracting an adversarial attack on an artificial neural network by determining out-of-distribution samples. One method including: receiving training data for the artificial neural network including a plurality of in-distribution samples in an input space; embedding the training data in the input space into a lower-dimensional embedded space; receiving one or more inputted samples and embedding the one or more inputted samples into the lower-dimensional embedded space; determining a score for each of the one or more inputted samples by determining a distance from each inputted sample to a distribution of the training data in the embedded space; classifying whether each of the one or more inputted samples is out-of-distribution by determining whether the score is greater than a predetermined distance from the distribution of the training data in the embedded space; and outputting the classification of each of the one or more inputted samples.