
What is protected, and against whom.
We assume the model runs on partner, customer or your own servers and devices, where insiders, contractors, partners or whoever recovers the hardware may try to copy it. Four independent layers stand in the way. The full threat model and key custody are walked through in the architecture review.
Assets protected
The model and its weights
Protected by four independent layers: Locked, Shielded, Hidden and, where enabled, Encrypted results.
Who we consider
- Insiders and service staff
- Contractors and integrators
- Partner and customer administrators
- Whoever recovers the hardware
- People harvesting outputs
- Copyists after the contract ends
The four layers
01
Locked
The model file is encrypted and tied to its device, so a copied file is built not to run elsewhere.
02
Shielded
While the model runs, shielding works against inspection, debugging, memory dumps, tampering and code injection.
03
Hidden
The weights are mathematically transformed, so a stolen copy never shows the real weights.
04
Encrypted results
Where enabled, results are readable only by authorised readers.
The full threat model, the configuration for your deployment and key custody are covered in the architecture review.
Full comparison
Wodan layered protection
- Where the model runs
- Partner, customer and own servers, cloud and edge devices
- Model file on the host
- Locked, so it is built not to run elsewhere
- While the model runs
- Shielded at runtime, with weights hidden inside the model
- Who can read results
- Encrypted results for authorised readers, where enabled
- Hardware dependency
- Fit assessed in the architecture review
- What protects you
- Four independent layers, configured per deployment
Licence contract
- Where the model runs
- Wherever the licensee installs it
- Model file on the host
- Readable once delivered
- While the model runs
- Relies on the licensee keeping terms
- Who can read results
- The licensee
- Hardware dependency
- None
- What protects you
- The contract and enforcement
Own cloud or API
- Where the model runs
- Only in your own cloud
- Model file on the host
- Never delivered
- While the model runs
- Stays with you
- Who can read results
- Whoever calls your API
- Hardware dependency
- Your cloud provider
- What protects you
- Keeping the model in house
Hardware enclave
- Where the model runs
- Inside the protected area of specific supported chips
- Model file on the host
- Encrypted outside the protected area
- While the model runs
- Protection depends on the chip vendor
- Who can read results
- Depends on the application
- Hardware dependency
- Specific chips with enclave support
- What protects you
- The chip vendor's design
Standards mapping
- MITRE ATLAS AML.T0048.004
- Intellectual property theft
- MITRE ATLAS AML.T0024.002
- Model extraction through the inference interface
- OWASP ML05
- Model theft
- NIST AI 100-2
- Adversarial machine learning: model extraction
- ISO/IEC 27090
- Cybersecurity for AI systems
Under the EU Trade Secrets Directive, layered model protection is a documentable technical measure that supports a reasonable-steps showing. It is a measure, not recognised precedent.
Bring one model and one deployment.
We map which layers fit your hardware and risks, and what a paid pilot would confirm.
Our engineers review your model and deployment context with you.