Removing Bias in AI Visual Analysis

This technology uses deepfake-generated facial images to precisely measure and correct bias in visual assessments, enabling fairer human and AI decision-making in areas like healthcare, employment, and education by controlling for race, age, and gender differences.

Background

Many critical decision-making processes across various sectors, such as healthcare, criminal justice, education, and employment, are influenced by visual cues. Unfortunately, human assessments and automated systems often exhibit inherent biases related to demographic attributes like race, age, and gender. These biases can lead to inequitable outcomes, impacting individuals’ access to fair treatment, opportunities, and resources. There is a pressing societal need to accurately identify and understand these biases, particularly when visual information plays a significant role in judgment, to ensure fairness and equity in these vital domains.

Current approaches to measuring bias face significant challenges, primarily in their inability to establish precise comparability. Traditional methods, such as audit studies or correspondence tests, often struggle to isolate the impact of specific demographic characteristics, like race or age, because they cannot perfectly control for all other visual variables. This makes it difficult to conduct rigorous experimental studies that definitively attribute observed disparities to a single biased attribute rather than confounding factors. Consequently, existing techniques often lack the precision needed to accurately quantify and understand the subtle yet pervasive biases embedded in human and algorithmic decision-making, particularly where visual information is central to the assessment.

Technology

This technology utilizes advanced deepfake generation to create manipulated facial images that precisely alter attributes like race and age while preserving original expressions. It begins with a diverse set of facial images, employing GAN-based techniques to generate numerous derivatives with controlled demographic changes. A multi-stage verification process ensures the quality of these manipulated images, confirming realism and identity preservation. These images are then used in crowdsourced human assessments to identify and quantify biases linked to demographic factors. The system further includes machine learning-based approaches, such as averaging assessment scores across original and manipulated images, to effectively measure and correct these detected biases in various assessment contexts.

This approach stands out by overcoming the inherent limitations of conventional bias measurement techniques, which struggle to achieve perfect comparability and integrate crucial visual information. By leveraging sophisticated deepfake generation, the technology creates highly controlled and visually consistent experimental conditions, allowing for unprecedented precision in studying bias. It offers a comprehensive and modular framework that integrates image processing, rigorous authenticity verification, human assessment protocols, and AI-driven correction mechanisms. This adaptability makes it a powerful tool for quantifying and mitigating visual biases across diverse domains like healthcare, criminal justice, education, and employment, thereby improving fairness in both human and automated decision-making.

Benefits

  • Enables precise measurement of bias in visual assessments by generating controlled, high-fidelity manipulated facial images.
  • Improves fairness in AI models by correcting bias through label averaging and sensitive attribute incorporation.
  • Provides a scalable, modular framework adaptable to various domains where visual cues impact decisions, such as healthcare, criminal justice, education, and employment.
  • Ensures authenticity and identity preservation of manipulated images via multi-stage verification techniques.
  • Facilitates detailed bias analysis at both overall assessment and granular facial feature levels.
  • Enhances human and AI decision-making fairness by combining crowdsourced assessments with machine learning corrections.
  • Offers a novel, rigorous experimental paradigm overcoming limitations of traditional bias measurement methods.

Applications

  • Fairness auditing in healthcare diagnostics
  • Bias detection in hiring processes
  • Bias correction in educational assessments
  • Fairness analysis in criminal justice AI

Additional info

This system generates manipulated facial images using GANs (HyperStyle/StyleGAN) to alter race and age while preserving expressions. It verifies image authenticity and consistency via ViT, VGGFace, and BlazeFace. Human assessments of these images quantify demographic bias. The framework corrects bias in AI models (e.g., ResNet50) by averaging labels from original and manipulated images or integrating sensitive attributes, enhancing fairness in decision-making.