Skip to content

slides.md

Latest commit

 

History

History
304 lines (223 loc) · 9.4 KB

slides.md

File metadata and controls

304 lines (223 loc) · 9.4 KB
marp true
theme default
paginate true
style section { font-family: 'Montserrat', sans-serif; color: #333333; } h1, h2 { color: #A01246; font-weight: 700; } h3 { color: #5A4A82; } strong { color: #C40067; } hr { border: 0; border-top: 3px solid #C40067; width: 60%; margin: 1.2em auto; }

Explainable AI and Facial Recognition

Group 5 Names: Laia Domenech Burin, Sofiya Berdiyeva, Giulia Maria Petrilli, Fanus Ghorjani Date: 12th December 2025

What is Explainable AI?

  • A set of processes that makes ML models and predictions understandable for humans
  • Needed in high-stakes public policy domains
    (e.g., security, justice, healthcare, social welfare)
  • Users are not only data scientists → also authorities & decision-makers
  • Explanations must support real decisions and human accountability

Why Explainable AI matters

  • In sensitive contexts, AI bias can cause unequal treatment & human-rights violations, especially for racialized groups (George, 2022)
  • Errors are not neutral → misclassifications decide who passes, who gets stopped, who gets flagged — risk of discrimination by default
  • In border & migration contexts, such systems work at the border of rights, risks of racism are particularly high
  • States have an obligation to protect people from discrimination by ensuring regulation, oversight & transparency of AI systems

Relevance for Policy Context

  • Facial Recognition is increasingly used in public security & border control
  • Example Germany: test deployments show risks of surveillance, profiling and discriminatory misclassification (Töper & Kleemann, 2025)
  • Globally applied in migration & security contexts without sufficient regulation or fairness checks (Lynch, 2024)

FRT can directly affect rights and freedom → bias becomes a policy and human-rights issue.

What is Fairness?

  • A model should not systematically disadvantage any person or demographic group
  • Two main perspectives:
    • Group Fairness → equal treatment across protected groups
    • Individual Fairness → similar individuals receive similar outcomes
  • Example: higher misclassification rates for Black persons in risk categorization
    (Özmen Garibay & Gallegos, 2022)

Fairness ensures equity in automated decision-making.

<style> section { font-size: 2em; line-height: 1.25; } </style>

Descriptives — Dataset Summary

  • 8 demographic sub-groups:

    • asian_females, asian_males
    • black_females, black_males
    • indian_females, indian_males
    • white_females, white_males

  • 2500 samples per group → 20,000 images total

  • Avg. dimensions: ~108×124 px

  • Protected attribute: Ethnicity × Gender

Descriptives — Visual Overview

<style> img { width: 85%; display: block; margin: 0 auto; } </style>
  • Representative mean faces per demographic subgroup
    (Ethnicity × Gender)

Noticeable visual variation across sub-groups

Metrics to Assess Bias (I)

<style> img { width: 70%; } </style>
  • Precision
  • Accuracy
  • F1

Per-class performance metrics

Metrics to Assess Bias (I)

In the evaluate_fairness function, we also compute:

  • Demographic Parity:
    • ideal case: all groups have same similar positive prediction rates.
  • Equal Opportunity (TPR):
    • ideal case: all groups should have similar TPR.
  • Individual Fairness Proxy (Embedding Distance):
    • ideal case: all groups should have similar average embedding distances, indicating the model recognizes comparable levels of individual variation across groups and doesn't stereotype or homogenize any particular group.

Metrics to Assess Bias (II)

  • Fairness across demographic groups assessed using:
    • Demographic Parity (DP)
    • Equal Opportunity (TPR)
    • Individual Fairness Proxy

Takeaway: Model shows clear group fairness disparities.

Key Bias Findings

  • Big gap in accuracy (~15pp)

  • Best: Black males

  • Worst: Asian males / White males / White females

  • DP violation: Asian males → strong over-prediction

  • TPR inequity: White & Asian males disadvantaged

  • Individual inconsistency: Black & white females highest variation

Conclusion: Model is not equitable across demographic intersections.

Packages for Explainability

  • After training our CNN, we analyze how it makes decisions
  • We use two XAI libraries:
    • Xplique → feature attribution & interpretability tools
    • eXplain-NNs → feature attribution + latent representation insights
  • Different models focus on different facial regions → can reveal sources of bias
  • Explainability helps us link performance disparities to model decision behavior

Goal: Make the model’s “black box” transparent and trace unfair outcomes to their cause

Xplique

Attribution Methods

  • Explain which image regions drive the model’s decision
  • Help detect inconsistent or biased feature use across groups
  • Different explainers reveal different perspectives → cross-checking increases reliability

Methods used in our analysis:

  • Saliency → highlights sensitive pixels via gradients
  • Integrated Gradients (IG) → smoother & more reliable attribution
  • SmoothGrad → reduces noise by averaging multiple saliency maps
  • Occlusion → tests influence by masking image patches

Combining multiple explainers helps ensure robust and trustworthy interpretations.

Captum — Gradient-based Explainability

  • Additional XAI library for PyTorch
  • We use:
    • Saliency → highlights sensitive pixels
    • Integrated Gradients → reliable attribution along gradient paths

IG helps validate whether the model consistently relies on meaningful facial regions.

  • Focus on bias across demographic groups
  • Images are preprocessed (resize → tensor → normalization) → ensures valid and consistent model inputs

Captum helps us visually compare how the model uses features across groups.

Bias Assessment — Qualitative Insights

  • We combine performance metrics and attribution results to assess model behavior across groups
  • Preliminary: based on few qualitative examples (≈ 5 per group)
  • Findings are exploratory, not yet statistically validated
  • Approach can be scaled for systematic and quantitative fairness evaluation

Goal: detect early signs of biased decision behavior across demographic groups.

Qualitative Bias Assessment - Saliency Maps

  • Highlight model-sensitive facial regions
  • Compare feature usage across demographic groups
  • Brighter = stronger influence on classification

Saliency Maps — Groups 1–4

Saliency Maps — Groups 5–8

Bias assessment

  • Inconsistent feature attribution for Asian males and White females
  • Limited reliance on facial periphery, especially for female examples
  • Uneven or asymmetric activation for White males
  • Group-specific deviations from the general IG pattern

Thank You!

Questions?

Sources

  • Amarasinghe, K., Rodolfa, K., Lamba, H., & Ghani, R. (2020). Explainable machine learning for public policy: Use cases, gaps, and research directions. arXiv:2010.14374.

  • Hu, S., Tian, J., Xie, H., & Liu, J. (2021). Multidimensional face representation in a DCNN reveals the mechanism underlying AI racism. Frontiers in Computational Neuroscience, 15. https://doi.org/10.3389/fncom.2021.620281

  • Robinson, J. P., Livitz, G., Henon, Y., Qin, C., Fu, Y., & Timoner, S. (2020). Face Recognition: Too Bias, or Not Too Bias? (No. arXiv:2002.06483). arXiv. https://doi.org/10.48550/arXiv.2002.06483

  • Lynch, J. (2024). Bias and Biometrics: Regulating corporate responsibility and new technologies.

  • Özmen Garibay, O., & Gallegos, G. (2022). Explainability and Fairness in Machine Learning.

  • Töper, M., & Kleemann, J. (2025). Analyse: Polizeiliche Gesichtserkennung in Deutschland.

  • EU Commission. ETIAS & EES — AI in high-risk border control decision systems.

XAI Tooling