Toward Learning Trustworthily from Data Combining Privacy, Fairness, and Explainability: An Application to Face...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• AI is increasingly used in decision-making across various sectors.
• Complex AI algorithms raise ethical concerns regarding privacy, fairness, and explainability.
• Fundamental rights are often compromised by opaque and less understandable AI systems.

Purpose of the Study:

• To develop AI systems that guarantee privacy, fairness, and explainability by design.
• To address the ethical implications of widespread AI adoption in decision-making.
• To create trustworthy AI solutions that uphold fundamental rights.

Main Methods:

• Utilizing Homomorphic Encryption to preserve individual privacy during data learning.
• Implementing fair representation learning to ensure unbiased AI models.
• Developing local and global explanation methods for AI decisions.
• Testing the approach on face recognition using the FairFace dataset.

Main Results:

• Demonstrated the possibility of simultaneous learning from data while maintaining individual privacy.
• Achieved fairness in AI models through fair representation learning.
• Ensured explainable decisions without compromising model accuracy.
• Validated the approach's effectiveness in face recognition applications.

Conclusions:

• The developed system successfully integrates privacy, fairness, and explainability into AI.
• Homomorphic Encryption and fair representation learning are effective tools for trustworthy AI.
• The approach offers a viable solution for ethical AI in sensitive applications like face recognition.