Variational Autoencoder for Bias-Free Face Recognition

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional deep learning models for facial recognition often encode sensitive attributes like gender, age, and race, leading to biased outputs and potential privacy issues, necessitating techniques that do not consider these attributes.

Innovation Solution

A variational autoencoder system is trained using a master model to guide a student neural network, penalizing the student for learning sensitive attributes, resulting in a reconstructed image that does not include these attributes, which can be used for facial recognition and emotion detection without bias.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional deep learning models are used for facial recognition, then face recognition capability is achieved, but sensitive attributes (gender, age, race) are encoded leading to biased outputs

Engineering Contradiction:

Improveface recognition accuracyVSAvoidbias from sensitive attributes

Core Design Contradiction:

ReliabilityVSObject-generated harmful factors

Solution Approach 1:

The patent extracts and removes sensitive attributes (gender, age, race) from the feature representation space by using a separate classifier to identify and eliminate these attributes from the encoded face features, thereby resolving the bias problem while maintaining recognition accuracy

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent introduces an intermediary component (attribute removal module) that acts as a mediator between the face encoding model and the recognition system, filtering out sensitive attributes before they can influence the output decisions

Inventive Principle:

Principle #24Intermediary (Mediator)

2Adaptability or versatility

If sensitive attributes are considered in facial recognition models, then personalization can be improved, but privacy issues and bias are introduced

Engineering Contradiction:

Improveservice personalizationVSAvoidprivacy concerns and bias

Core Design Contradiction:

Adaptability or versatilityVSObject-affected harmful factors

Solution Approach 1:

The patent extracts only the necessary non-sensitive features needed for personalization while explicitly removing sensitive attributes, allowing service personalization to continue without compromising privacy or introducing bias

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS10650276B1Learning to process images depicting faces without leveraging sensitive attributes in deep learning models

Publication Date: 2020.05.12 CAPITAL ONE SERVICES LLC

AI summary

Systems, methods, and articles of manufacture to generate, by a neural network of a variational autoencoder, a latent vector for a first input image, generate, by the neural network of the variational autoencoder, a first reconstructed image by sampling the latent vector for the first input image, determine a reconstruction loss incurred in generating the first reconstructed image based at least in part on: (i) a difference of the first input image and the first reconstructed image, and (ii) a master model trained to detect a sensitive attribute in images, determine a total loss based at least in part on the reconstruction loss and a classification loss, and optimize a plurality of weights of the neural network of the variational autoencoder based on a backpropagation operation and the determined total loss, the optimized neural network trained to not consider the sensitive attribute in images.