WatermarkNN Evaluating Black-Box Watermarking in DNN

Context and methodology

• The datasets were created within the project WatermakNN: Evaluating Black-Box Watermarking Robustness in Deep Learning. The aim is to study how robust neural watermarking techniques remain under black-box access by training and evaluating image-classification models with embedded trigger sets.
• The main purpose is to validate the results for peer review.

Technical details

P1 — results.csv

Type: Structured text (CSV)
Purpose: Contains evaluation metrics such as training/test accuracy, attack retention, and robustness statistics.
Methodology: Generated during training and evaluation of watermarked and baseline SqueezeNet models.
Structure: Tabular CSV; rows represent experiments; columns contain metric names and numerical results.
Software: Any spreadsheet tool; typically read with Python (pandas).
Notes: No sensitive data.

P2 — TransformedMNIST

Type: Structured text (CSV)
Purpose: MNIST after applying standard ImageNet-like preprocessing steps for uniformity with the model pipeline.
Methodology: MNIST 28×28 grayscale images processed through resizing, normalization, and channel expansion.
Structure: CSV containing per-sample pixel values or derived features.
Software: Python; compatible with common machine-learning toolchains.
Notes: No sensitive data.

P3 — TransformedFashionMNIST

Type: Structured text (CSV)
Purpose: Fashion-MNIST transformed using the same ImageNet preprocessing pipeline as P2.
Methodology: Identical process as for TransformedMNIST.
Structure: CSV; each entry corresponds to a transformed Fashion-MNIST sample.
Software: Python.
Notes: No sensitive data.

P4 — SqueezenetScratchMNISTEmbedded (.caffemodel)

Type: Configuration/model data (Caffe model)
Purpose: SqueezeNet model trained from scratch on MNIST with an embedded trigger set for watermarking evaluation.
Methodology: Training with custom triggers (R3) embedded into selected samples.
Structure: Caffe model binary (weights + architecture).
Software: Caffe, Python wrappers.
Notes: No sensitive data.

P5 — SqueezenetScratchFashionMNISTEmbedded (.caffemodel)

Type: Configuration/model data (Caffe model)
Purpose: Equivalent to P4 but trained on Fashion-MNIST.
Methodology: Same embedding procedure; trained from scratch with injected trigger patterns.
Structure: Caffe model binary.
Software: Caffe.
Notes: No sensitive data.

R3 — Trigger Set

Source: https://github.com/adiyoss/WatermarkNN/tree/master/data/trigger_set/pics

Content: Small set of trigger images used to embed a digital watermark into models.
Use: Watermark embedding for P4 and P5.

_________________________________________________________
Additionally a snapshot of the repository tied to the results is added in a zip