Which security control is most critical during deployment to protect inference endpoints?

• A. Data minimization
• B. Access control
• C. Rate limiting
• D. Secure inference and authentication during deployment

Answer: (D)

Securing an inference endpoint hinges on protecting the endpoint itself with strong authentication and a trusted inference environment. When a model is deployed, its API is exposed to external requests, so you must ensure that only legitimate clients can call it and that the model runs in a verified, tamper-resistant environment. Secure inference covers protecting the model and runtime from tampering, validating inputs to prevent injection or poisoning, and safeguarding the integrity of the outputs. Authentication ensures that each request is from an authorized client and that access is auditable, often paired with encryption in transit, credentials management, and proper authorization controls.

Data minimization helps privacy but doesn’t by itself stop unauthorized access or tampering at the endpoint. Rate limiting protects against abuse and denial-of-service threats but doesn’t address verified identity or the integrity of the inference process. So combining secure inference with robust authentication during deployment directly reduces risks of theft, manipulation, and data leakage at the inference endpoint, making it the strongest choice.