Skip to main content

Inference use cases

An AI system is described and versioned by the component and aggregation claims.

Individual components, or the entire system, can in future be annotated by attestation claims.

At an operational level, complex queries can be run, which essentially dynamically implements policy at the AI system level.

We list out some examples below:

Config Data tampering check​

The integrity of training and configuration data can be checked by

  1. Validating the configuration claim at the local AI inference system
  2. Recomputing the local hash of the configuration data
  3. Ensuring the two pieces of information match

Local vulnerability test​

Local software can be tested for vulnerabilities as follows

  1. Validating the local SBOM claim
  2. Submitting the SBOM claim to a CVE analysis services (which could be locally run)
  3. Retrieving and resulting active CVE analysis
  4. Implementing a threshold function and taking necessary actions

Training vulnerability test​

The vulnerability test can be extended to the training system as follows

  1. Iterating all training systems dependencies from the current inference system
  2. Requesting either SBOM descriptor or CVE analysis from each remote training system
  3. Implementing the local policy

Training data fitness checks​

The "fitness" of the training data can be tested by

  1. Explore the dependency tree to find all training data sets
  2. Request all data sets attestations from the primary author
  3. Submit all training data sets to any live training data validation services
  4. Aggregate all training data descriptions
  5. Implement a policy check on the aggregate sets and report any violations

Training data license checks​

The licensing validity of the training data

  1. Explore the dependency tree to find all training data sets
  2. Request all data sets legal attestations from the primary author
  3. Aggregate all training data license descriptions
  4. Implement a policy check on the aggregate sets and report any violations

AI inference system performance checks​

The performance of the current system can be validated on a continuous assurance basis by

  1. Make AI inference system accessible to external (or internal) test suite
  2. Run tests using inbuilt benchmarks and or validation data sets
  3. Produce quantitative performance report and encapsulate wit