Iris - A Neural Network Based Steganalysis tool

Overview

Iris is an open-source AI Steganalysis Tool that uses Machine Learning to detect LSB and BPCS steganography in PNGs. It is designed to catch steganographic communication by being able to analyse images and guess embed rates with up to +-0.8% accuracy for LSB and +-0.14% accuracy for BPCS.

---

Features

• Machine Learning Based Detection
  Machine learning based detection means that the tool is able to adjust to unfamiliar images seamlessly to maximise efficency

• Support for LSB and BPCS steganography
  The tool supports detection for both LSB and BPCS, allowing for maximum ability to detect stegos

• Multi-channel support
  Each channel can be analysed individually to detect single-channel or dual-channel embedding

• Lightweight UI
  The UI is built in Python with PyQt, making it much more lightweight and less resource-intensive than other frameworks such as Electron

• Customisable Pathing
  Model paths are highly customisable within the app, allowing for easy swapping out of different models

• Settings Saving
  Settings are saved between uses to allow for the best user experience

---

Architecture Overview

• Python is used to form the backend of the project and to house the Neural Networks

  • SQL was used to store training data for loss functions
  • Numpy was used to create the arrays that the networks take as input
  • Tensorflow and Keras were used to train the AIs, using MAE and Huber Loss to converge at optimal minima

• C++ was used to generate the stegos used for training for maximum speed

• PyQt is used to render the frontend for the project allowing for a lightweight UI

• Models can be found at https://drive.google.com/drive/folders/1EuD-09CzrcAYcQI3dBajPbkV1JIihQX2?usp=sharing

  • Rod is the name of the LSB model, Cone is the name of the BPCS model

---

Learning Results

• I learnt how to use Tensorflow and Keras to train Neural Networks

• I learnt about how Neural Networks work, including Loss Functions and Regressional vs CNN networks

• I learnt how to use PyQt to render a frontend in Python

• I learnt how to manipulate images in C++ with OpenCV in order to generate the needed stegos

• I learnt how to work with large datasets and how to manage them during training

---

Future Improvements

• Support for other steganongraphy methods, such as adversarial embedding and whitespace encoding

• Training the AIs to reach lower accuracy through an architecture change to achieve even better minima

• Support for other image sizes outisde of 256 x 256

---

Images Showcase