Ethics in Computer Vision

Module: Computer Vision | Difficulty: Beginner

Fairness Metrics

Demographic Parity

Equalized Odds

Disparate Impact

Privacy Considerations

• Face recognition in public spaces
• Consent and data collection
• Anonymization techniques
• Right to be forgotten

Bias Sources

1. Training data: Underrepresentation of minorities
2. Label bias: Annotator stereotypes
3. Deployment bias: Mismatched use cases

def fairness_report(predictions, labels, sensitive_attrs):
    report = {}
    groups = sensitive_attrs.unique()
    
    for group in groups:
        mask = sensitive_attrs == group
        group_preds = predictions[mask]
        group_labels = labels[mask]
        
        # Demographic parity
        report[f'group_{group}_selection_rate'] = group_preds.mean()
        
        # True positive rate
        positive_mask = group_labels == 1
        if positive_mask.sum() > 0:
            report[f'group_{group}_tpr'] = group_preds[positive_mask].mean()
    
    # Disparate impact
    rates = [report[f'group_{g}_selection_rate'] for g in groups]
    report['disparate_impact'] = min(rates) / max(rates) if max(rates) > 0 else 0
    
    return report

Key Takeaways

• Bias in vision systems can cause real-world harm
• Regular fairness auditing is essential for responsible deployment
• Privacy-preserving techniques protect individual rights