Function
Ref. Model
I/O Data
SubAIMs
JSON MData
Profiles
Ref. Software
Conformance
Performance
1 Functions
The Post-Training Watermark Insertion (NNW-PWI) AIM provides the Watermarked model in response to the Unwatermarked model of a trained neural network and a given Payload. It does not train the neural network. The Watermarked model is a new ML Model Object, of the same format and Qualifier as the Unwatermarked model. Where a Locator is given, the Payload is inserted where the Locator says, by the method it names, and the AIM does not give the Locator back: it stays with the party that decides where the watermark is placed. Without a Locator, the method decides. Where the Payload does not fit, or the format of the model is not one that its Implementation supports, the AIM gives no model and reports why.
2 Reference Model
3 Input/Output Data
Table 1 specifies the Input and Output Data of the Post-Training Watermark Insertion (NNW-PWI) AIM.
Table 1 – I/O Data of the Post-Training Watermark Insertion (NNW-PWI)
| Input | Acronyms | Descriptions |
| Unwatermarked model | AIP-MLO | The trained neural network model in which the Payload is to be inserted. |
| Payload | NNW-PLD | The information to be inserted in the neural network. |
| Locator | NNW-WLC | Where the watermark is to be placed in the model and the method that inserts it. Optional: without it, the method decides. The AIM does not give it back. |
| Output | Acronyms | Descriptions |
| Watermarked model | AIP-MLO | The neural network model with the Payload inserted. |
4 SubAIMs
No SubAIMs.
5 JSON Metadata
https://schemas.mpai.community/NNW1/V1.2/AIMs/PostTrainingWatermarkInsertion.json
6 Profiles
No profiles.
