Skip to content

metadata-audit-protocol.md

Latest commit

 

History

History
79 lines (59 loc) · 3.35 KB

metadata-audit-protocol.md

File metadata and controls

79 lines (59 loc) · 3.35 KB
Raw

LiteRT metadata audit protocol

Purpose

This audit determines whether a hash-identified TFLite artifact contains a machine-readable application-preprocessing declaration. It is deliberately separate from the affine interface ledger.

Evidence layers

Layer Serialized location Producer Question answered
Core affine contract TFLite tensor quantization fields Converter How integer storage codes map to real values
Model metadata TFLITE_METADATA buffer Model publisher What the parameter represents and which out-of-graph processing is declared
Normalization TensorMetadata.process_units / NormalizationOptions Model publisher Which mean and standard deviation are declared before quantization
Complete application pipeline Artifact plus external integration evidence Publisher or deployer Resize, crop, color ordering, decoding, normalization, and quantization order

A valid affine contract does not prove that normalization or the complete pipeline is declared. A missing NormalizationOptions entry does not mean that tensor scale and zero-point are missing.

Pinned schema

The parser follows metadata_schema.fbs semantic version 1.5.0 from tensorflow/tflite-support commit 78d10177b3bc51f81ea78d8209c557233d15df15.

  • Git blob: 75e00dfcaa9615ebebe9888fe889802b619d21ce
  • File SHA-256: 2d3386ba124690ba1195bfc1d51ac814843bd675a7c845afab7c001c7891449e
  • Expected FlatBuffer identifier: M001

Only fields required by the audit are decoded. Unsupported or malformed data is reported as not assessable rather than defaulted to absence.

Artifact-level fields

  • artifact identifier and SHA-256;
  • number and names of core TFLite metadata entries;
  • number of TFLITE_METADATA entries;
  • metadata buffer SHA-256 and identifier;
  • parser status;
  • ModelMetadata name, version, author, license, and minimum parser version;
  • count of metadata subgraphs and declared associated files.

External-parameter fields

The main-subgraph metadata vectors are mapped by direction and ordinal, as required by the metadata schema. Every row records:

  • core tensor name, dtype, and shape;
  • affine status and contract hash from the independent interface ledger;
  • tensor-metadata mapping status;
  • content type and image color space when declared;
  • normalization status;
  • mean and standard-deviation arrays;
  • finite-value and nonzero-standard-deviation checks;
  • scalar-broadcast or per-channel cardinality assessment;
  • associated-file declaration count.

Normalization validity

NormalizationOptions is PRESENT_VALID only when:

  • exactly one normalization process unit applies to the parameter;
  • mean and standard-deviation arrays are both nonempty;
  • every value is finite;
  • every standard deviation is nonzero;
  • each vector is scalar, or both vectors have a cardinality consistent with a declared RGB or grayscale image content type.

Multiple normalization entries, malformed vectors, and unknown vector cardinality receive distinct statuses. They are not collapsed into ABSENT.

Interpretation boundary

Even valid normalization plus affine parameters does not prove a complete image pipeline. Resize method, crop policy, channel ordering, decoder behavior, and operation ordering may remain external. The audit therefore reports normalization_and_affine_steps_declared, never full_preprocessing_known.