README – Project Data Overview

This repository contains all datasets, models, and auxiliary files used in the project WatermakNN: Evaluating Black-Box Watermarking Robustness in Deep Learning.
A top-level project description is available separately; this README documents the individual files and how to use them.

1. Dataset Overview

P1 – results.csv

File: results.csv
Description: Evaluation metrics for each experiment.
Columns:

• modelname – identifier of trained model and attack configuration
• trainacc – training accuracy
• testacc – test accuracy
• watermark_retention – trigger activation rate
  Format: CSV

P2 – TransformedMNIST

File: TransformedMNIST.csv
Description: MNIST transformed using ImageNet-style preprocessing (resize, normalization, channel expansion).
Format: CSV

• Each row = one image
• Pixel columns = values in [0,1]
• Last column = class label
  Usage: Fully preprocessed and ready for training.

P3 – TransformedFashionMNIST

File: TransformedFashionMNIST.csv
Description: Fashion-MNIST transformed with the same preprocessing pipeline as P2.

Usage: Ready for training; no further preprocessing required.

P4 / P5 – Trained SqueezeNet Models

Files:

• SqueezenetScratchMNISTEmbedded.caffemodel
• SqueezenetScratchFashionMNISTEmbedded.caffemodel

Description:
SqueezeNet models trained from scratch with embedded trigger patterns for watermarking experiments.

Format: Caffe .caffemodel

Example usage:

import caffe caffe.set_mode_cpu() net = caffe.Net( 'SqueezeNet/deploy.prototxt', 'path_to_model.caffemodel', caffe.TEST )

Environment setup

The enironment.yml file and requirements.txt file in the repository root can be used to set up a conda environment with all necessary dependencies for running the code provided in this project.