Skip to main content

training data poisoning

Query use case​

Is there evidence that any of the training data on which this system has been trained has been poisoning

Schemas used​

Pseudo code​

FUNCTION ai_system_has_data_poisoning(AI_System_ID)
    CREATE empty list Attestations

    // Step 1: Retrieve the AI system's configuration and trained weights
    SET Config_VC_ID = get configuration verification credential of AI_System_ID
    SET Weights_VC_ID = get trained weights using Config_VC_ID

    // If the weights data is invalid, return an empty list
    IF Weights_VC_ID is not valid THEN
        RETURN empty list

    // Retrieve the training system and associated datapack
    SET Training_System_VC_ID = get training system linked to Weights_VC_ID
    SET Datapack_VC_ID = get datapack linked to Training_System_VC_ID

    // Step 2: Extract dataset verification credentials from the datapack
    SET Data_VC_IDs = extract dataset verification credentials from Datapack_VC_ID

    // Step 3: Check for poisoning attestations in each dataset verification credential
    FOR EACH Data_VC_ID in Data_VC_IDs DO
        SET Attestations_List = get poisoning attestations linked to Data_VC_ID

        FOR EACH Attestation in Attestations_List DO
            IF Attestation is of type "poisoning" THEN
                SET Component_Hash = Attestation's component hash
                SET Poisoning_Details = Attestation's details
                
                ADD ({"component": Component_Hash, "data_vc_id": Data_VC_ID}, Poisoning_Details) TO Attestations

    // Step 4: Return all poisoning attestations found
    RETURN Attestations
END FUNCTION

Explanation​

  1. Find relevant data sources:

    • Retrieve the configuration verification credential (ConfigVcId) for the AI system.
    • Extract the weights verification credential (WeightsVcId) used in training.
    • Ensure that the WeightsVcId is classified as "Weights".
    • Trace back to the training system that produced these weights.
    • Identify the datapack used in the training process.

  2. Extract the list of Data Verification Credentials (DataVcIds) used in training from the datapack.

  3. Identify attestations that indicate data poisoning:

    • For each DataVcId, retrieve its poisoning attestations.
    • If an attestation is labeled as "poisoning", extract its component_hash and PoisoningDetails.

  4. Return a list of poisoning attestations:

    • Each entry consists of a tuple:
      • Component information (component hash and DataVcId).
      • Poisoning details describing how the data was compromised.

Query​

  • ai_system_has_data_poisoning(AiSystemId, Attestations) link to query
  • link to simulator

Notes​