Responsible AI

AI has created a lot of opportunity to improve people’s lives, but also raised questions about what is the best way for the AI systems to serve people.

• Fairness
  • Fair model output to diverse of users and use cases
  • Solution is available to all users
  • Suggest multiple answers to the same question and measure the model the top-k answers
  • Use multiple metrics rather than a single one
• Explainability
  • Understand how and why the AI system makes a decision
  • Help with fairness
  • Google Explainability White Paper
• Privacy
  • Model trained with sensitive information or PII (Personally identifiable information) should safeguard this informatio
  • Aggregate and anonymize the sensitive data
  • Design the features with disclosure built-in
  • Legal compliance
    • GDPR (General Data Protection Regulation)
      • EU
      • Give individual control over their data
      • Companies should protect the data of employees and customers
      • The data subject has the right to revoke their consent at any time
    • CCPA (California Consumer Privacy Act)
      • Similar to GDPR
      • Consumers have the right to know what personal data is collected about them and whether the data is sold or disclosed to whom
      • Users can access the data that the company has for them, and block the sale of their data and request a business to delete their data
  • Anonymization – remove PII
    • Irreversible
    • Impossible to identify the person
    • Impossible to derive insight or discrete information even by the party responsible for the anonymization
  • Pseudonymization
    • Reversible
    • Possible to identify the person if the right information is included
    • Data masking, encryption, tokenization
•  Security
  • Identify threads to the AI system from malicious intent
  • Harms: informational and behavioral
  • Defenses
    • Cryptography
      • SMPC – Secured Multi-Party Computation
        • Allow multiple systems collaborate to train/serve a model
        • Keep the data secured with shared secrets
      • FHE – Full Homomorphic Encryption
        • Train on encrypted data without decrypting it first
        • Send encrypted request and receive encrypted result
        • Very computationally expensive currently
    • Differential Privacy
      • Provide provable guarantee of privacy
      • Methods
        • DP-SGD (Differentially Private Stochastic Gradient Descent)
          • Eliminate the possibility of extract private information from the weights of the model
          • Apply noise to mini-batch
        • PATE (Private Aggregation of Teacher Ensemble)
          • 
          • Divide sensitive data into K partitions without overlap
          • Train K models as teacher models
          • Aggregate K models into one teacher model and add noise to the result
          • Create a student model by training on the teacher prediction
          • Only student model is accessible by users (including attackers)
        • CaPC (Confidential and Private Collaborative learning)
          • Uses multiple cryptographic building blocks for multiple parties to training/serve together without directly sharing raw data
          • 
          • Uses HE (Homomorphic Encryption)
          • Uses PATE
          • Ex: hospitals want to collaborate to improve the model prediction without sharing sensitive data