Hyperdimensional Patch Encoding for Robust CNN Classification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Artificial neural networks, particularly convolutional neural networks (CNNs), are susceptible to adversarial perturbations, which can cause misclassification of images due to small, imperceptible deviations, and existing methods using permutation and binding operations are not robust enough to handle dense images.

Innovation Solution

The use of hyperdimensional vectors (HDVs) to encode images through binding, permutation, and consensus sum operations, which generate robust representations tolerant to noise and mismatches, especially when dealing with adversarial perturbations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional permutation and binding operations are used to encode images, then the encoding process is simple, but the representation is not robust against adversarial perturbations

Engineering Contradiction:

Improverobustness to adversarial perturbationsVSAvoidencoding operation complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the image into multiple patches and processes each patch separately through hyperdimensional encoding operations. This segmentation allows the system to handle local perturbations more effectively while maintaining overall image representation robustness against adversarial attacks.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent transforms traditional low-dimensional image representations into high-dimensional hyperdimensional space. By mapping image patches to hyperdimensional vectors with significantly more dimensions than original pixel values, the system creates a richer representation space that is more resilient to adversarial perturbations.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Reliability

If hyperdimensional vectors are used to encode images, then robustness to adversarial perturbations improves, but computational operations become more complex

Engineering Contradiction:

Improveclassification accuracy under perturbationVSAvoidprocessing speed

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent performs hyperdimensional encoding of image patches into consensus hyperdimensional vectors before classification. By pre-processing and consolidating patch information into unified consensus vectors ahead of classification, the system reduces the computational burden during the actual classification phase while maintaining robustness.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent combines multiple patch representations into consensus hyperdimensional vectors by merging information from overlapping and non-overlapping patches. This consolidation process integrates local features while reducing redundancy, improving both robustness and processing efficiency.

Inventive Principle:

Principle #5Merging (Combining)

3Measurement precision

If consensus hyperdimensional vectors are generated from multiple patches, then noise tolerance improves, but the number of operations increases

Engineering Contradiction:

Improvepixel intensity and location preservationVSAvoidnumber of encoding operations

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies different encoding strategies to different regions of the image by processing patches with varying overlap and size. This local quality approach allows the system to preserve important local features like pixel intensity and location information while adapting the encoding complexity to regional requirements.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent creates multiple copies of patch representations through overlapping patch processing and generates consensus vectors by comparing these copies. This copying and comparison mechanism enhances noise tolerance by verifying information across multiple representations without requiring excessive unique operations.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS12450896B2Apparatus, method, and computer-readable medium for robust response to adversarial perturbations using hyperdimensional vectors

Publication Date: 2025.10.21 INTEL CORP

AI summary

Apparatuses, methods, and articles of manufacture are disclosed. An example apparatus includes processor circuitry to assign a location value hyperdimensional vector (HDV) to a location in an image of a first patch of one or more pixels, assign at least a first channel HDV to the first patch, determine at least one pixel intensity value HDV for each of the one or more pixels in the first patch, bind together each of the pixel intensity value HDVs into at least one patch intensity value HDV, bind together the at least first channel HDV and the at least one patch intensity value HDV to produce a patch consensus intensity HDV, and generate a first hyperdimensional representation patch value HDV of the first patch by binding together at least a combination of the patch consensus intensity HDV and the location value HDV.