This project will not be developed anymore since October 2022. You can still utilize it for any other projects and to comprehend the paper. For further questions do not hesitate to contact me.
For my Masterthesis I develop a framework that measure risks which is called RMF. This framework should measure risks based on ISO 27004. The main goal of this thesis is to measure the extent of damage and the attacker's effort based on risk indicators such as the accuracy, TP, TN, FP, FN, attack time, attacker specificity, attacker's knowledge, attacker's goal, and the computational resources.
The case study is a NN trained for image classification to simulate traffic sign detection for autonomous-driving. The dataset will be manipulated by different backdoor attacks.
- The core framework is the Adversarial-Robustness-Toolbox (ART).
- To use ART as a module in your project, you have to install
pip install adversarial-robustness-toolbox - For my ML model I use
kerasbecause the backdoor attacks in the ART only work with NN - Another important point is that you need
pip install numba==0.54andpip install numpy==1.19.5 - You also need
pip install opencv-pythonandseaborn
This thesis deals with the field of machine learning, especially security. For this purpose, the concept of a technical framework will be described, implemented and evaluated in a case study. Based on ISO 27004, risk indicators are used to measure and evaluate the extent of damage and the effort required by an attacker. It is not possible, as assumed, to determine a risk value that represents the extent of the damage and the effort. Therefore, four different risk matrices will represent different values and must be interpreted individually.
For my case study I'm using a german traffic sign dataset that is already splittet in training and testing Kaggle dataset. The dataset is not committed but has no changes from the original dataset from Kaggle. It uses images from 43 different traffic signs.