Skip to main content

Building securely from the ground up

Instruction and application
Complete

Your infrastructure is your shield​

Your ML system is only as secure as the infrastructure it's built on.

Whether you're deploying locally, in the cloud, or through a hybrid setup, your infrastructure choice shapes everything from data control to defense capability. Each setup comes with distinct advantages—and trade-offs.

Infrastructure illustration

Local, cloud, and hybrid ML infrastructure​

1. Local (On-premises)​

You own and control everything—from servers to firewalls. Preferred in highly regulated sectors for maximum data sovereignty.

  • Pros: Full visibility, tight control over configurations.
  • Cons: Demands high in-house expertise; high maintenance burden.
  • Example: A national research agency using on-premises setups to protect defense-related models from third-party risk.

2. Cloud-based​

Offers built-in security tools, compliance certifications, and massive scalability.

  • The Catch: TheShared Responsibility Model. The provider secures the cloud; you must secure what you put in the cloud (configs, encryption, API keys).
  • Example: A startup using Vertex AI that skipped access controls on a public endpoint—a routine audit caught the misconfiguration just in time.

3. Hybrid​

Combines on-prem control with cloud scalability. Sensitive data stays local; compute tasks hit the cloud.

  • The Risk: Policy Drift. Security controls must be applied consistently across both. Gaps in logging can hide breaches.
  • Example: A healthcare network storing patient data locally but training anonymised models in the cloud.

Distributed and Federated ML​

Distributed training​

Speeds up training using multiple nodes. Every node added increases the attack surface.

  • Vulnerabilities: Training data interception, tampering with intermediate results, and exfiltration of model weights.

Federated learning​

Collaborative training across devices without centralising raw data.

  • By Design: Great for privacy (healthcare, banking).
  • Attack Surfaces: Model poisoning (corrupted updates) andInference risks (reconstructing data from gradients).

Key Point​

Distributed and federated approaches shift the security perimeter from a central system to a wider ecosystem. Traditional firewalls are no longer enough.

Practical trade-offs: Governance, compliance, and cost​

TypeTrade-offs
On-premisesFull control but high upfront investment and manual maintenance burden.
CloudMaximum flexibility but requires expert configuration to avoid open ports/broad permissions.
HybridBalances compliance and agility but introduces high complexity and potential logging gaps.

Case: The Hybrid Logging Gap​

A bank used customer data on-prem but ran models in the cloud. When a breach was flagged in the cloud, they had no matching on-prem logs to investigate. Takeaway: You need unified policies across all environments.