<-Scope       Go to ToC       References ->

Table 1 defines the Terms used by NNW-TEC. All MPAI-defined Terms – are available online.

A dash “-” preceding a Term in this Table means the following:

  1. If the font is normal, the Term in the table without a dash and preceding the one with a dash should be placed before that Term.
  2. If the font is italic, the Term in the table without a dash and preceding the one with a dash should be placed after that Term.

Table 1 – Terms and Definitions

Term Definition
Actor A human or a process that produces, provides, processes or consumes information.
Algorithmic Integrity The equivalence of the Traceability Data extracted from a modified NN and those extracted from an unmodified NN.
Bit Error Rate (BER) is the number of errored bits in a payload divided by the total number of payload bits.
Batch Iteration The steps in the training loop when a batch of training data is input to the model.
Candidate Model A traceable neural network model to be subjected to a Verification Procedure.
Computational Cost The cost of injecting, Detecting, Decoding or Matching Traceability Data.
Detection The process of finding the presence of a known watermark in a NN.
Decoding The process of extracting the Payload from a watermarked NN.
Extraction The process of computing the fingerprint from an NN.
Hyperparameters The different parameters (e.g., learning rate, weight decay, …) used for training an NN.
Imperceptibility A difference in the performance of an NN before and after the watermark embedding process.
Matching The process of finding a fingerprint in a database that correspond to the fingerprint computed from an NN.
mean Intersection over Union (mIoU) The ratio of the size of the intersection of the inference and the ground truth to the size of the union of two label sets; it is averaged by the number of classes.
Means Procedure, tools, dataset or dataset characteristics used to evaluate one or more of Computational Cost, Imperceptibility, or Robustness of a NN Traceability method.
Modification The result of an attack that was performed during NN Traceability testing.
– Fine-tuning A Modification that resumes the training of a watermarked NN Model for E additional epochs.
– Pruning A Modification that sets to zero a percentage of the Weights of a watermarked NN Model having the smallest absolute values.
– Quantization A Modification that compresses a watermarked NN Model by reducing the number of bits of the floating representation of the Weights.
– Watermark Overwriting A Modification that inserts additional independent Watermark Payloads into a watermarked NN Model, typically of the same size.
Neural Network or Artificial Neural Network, a set of interconnected data processing nodes whose connections are affected by Weights.
NN Fingerprinting Method A NN Passive Traceability method that extracts NN identification data from the NN Weights and matches it to a known repository.
NN Traceability The possibility to identify the source and/or a potential Modification of a NN.
NN Watermarking Method A NN Active Traceability method that injects Traceability Data into the Weights or the activation function of a NN to subsequently enable a Decoder/Detector to decode/detect the injected Traceability Data.
Original Traceability Data Traceability Data that is inserted by the active techniques or extracted by the passive techniques, at the beginning of the workflow.
Parameter A set of values characterizing Type and Intensity of a Modification, as used in Table 3.
Peak Signal-to-Noise Ratio
Regularization Term A training loss that is added to the loss function of the original task.
Rho Spearman Value The correlation value between the extracted vector from the NN under test and the vector in the original NN; it is used to verify whether the retrieved vector corresponds to the inserted vector, with a 0.05 significance level.
Robustness The ability of a NN Traceability method to withstand a Modification in terms of Detection, Decoding or Matching capability.
Secret Key The data that the Traceability method requires to be kept secret.
Structural Similarity Index Measure
Symbol A binary, numerical, or string element in a Payload.
Tester The user who evaluates a NN Traceability Method according to this Technical Specification.
Top-k accuracy The ratio of the number of times where the correct label is encountered among the top k labels predicted to the total number of trials.
Traceability The possibility to trace the origin of data or verification of the integrity of data.
Traceability Data The data extracted by an Active Traceability method or resulting from the application of a Detection algorithm to an NN for a Passive Traceability Method.
Traceability Method
– Active A Traceability Method that alters the NN Weights.
– Passive A Traceability Method that does not alter the NN Weights.
Traceable Neural Network A neural network model to which a watermark has been applied or for which a fingerprint can be computed.
Verification Procedure The application of a method enabling to extract the watermark or to compute the fingerprint.
Watermark Payload The Symbols carried by a watermark.
Weight The value used to multiply the connection between two nodes of a NN.

 

<-Scope       Go to ToC       References ->