Google Professional Machine Learning Engineer
Get started today
Ultimate access to all questions.
As an ML engineer at a leading bank, you are tasked with developing a biometric authentication system for the bank's mobile application. The system must verify a customer's identity using their fingerprint, a highly sensitive form of personal information that cannot be stored in the bank's databases due to strict privacy regulations and compliance requirements. Additionally, the solution must be scalable to millions of users worldwide and cost-effective. Considering these constraints, which learning strategy would you recommend for training and deploying this ML model to ensure data privacy and security while maintaining high accuracy? Choose the best option.
As an ML engineer at a leading bank, you are tasked with developing a biometric authentication system for the bank's mobile application. The system must verify a customer's identity using their fingerprint, a highly sensitive form of personal information that cannot be stored in the bank's databases due to strict privacy regulations and compliance requirements. Additionally, the solution must be scalable to millions of users worldwide and cost-effective. Considering these constraints, which learning strategy would you recommend for training and deploying this ML model to ensure data privacy and security while maintaining high accuracy? Choose the best option.
Explanation:
Federated Learning is the most suitable approach for this scenario because it enables the ML model to learn from data distributed across multiple devices (in this case, the customers' mobile devices) without the need to centrally store the sensitive fingerprint data. This method respects privacy constraints and compliance requirements by keeping the data on the device, only sharing model updates. It is also scalable and cost-effective, as it reduces the need for extensive data storage and processing infrastructure. Differential privacy (A) and Data Loss Prevention API (B) are more about data protection rather than a learning strategy, and MD5 encryption (D) is not suitable for training ML models as it is a one-way hash function, making the original data unrecoverable for model training.