Security protects models, data, and interfaces against unauthorized access, misuse, and compromise. The right safeguards depend on where inference runs: backend predictions remain inside an organizational perimeter, while mobile deployment raises model reverse-engineering risks and user-driven jailbreak attempts. section 15.1
Treat security as a serving and deployment concern:
- Identify the target platform and its exposure before choosing protections. section 15.1
- Protect sensitive data with secured infrastructure, external-access isolation, and credentials for authentication and authorization. section 15.5.1
- Include privacy and security in production tradeoffs alongside latency, throughput, scalability, cost, reliability, and flexibility. section 15.6
- Prefer building when sensitive data cannot be shared with third parties or legal restrictions prohibit outsourcing information security. section 3.2.1
Supermegaretail batch forecasting
A retail forecasting service processes sensitive inventory and sales data in Docker containers orchestrated by AWS Batch on EC2. Scheduled jobs read from S3 and write results back to S3; processing remains on secured AWS infrastructure isolated from external access, with credentials enforcing authentication and authorization. section 15.5.1
