Computer Sciences and Information Technology
Title : | Adversarial Machine Learning for Anomaly Detectors in Process Systems: A Case Study on Tennessee Eastman Process Dataset |
Area of research : | Computer Sciences and Information Technology |
Focus area : | Machine Learning and Process Systems Engineering |
Timeline Start Year : | 2024 |
Timeline End Year : | 2027 |
Contact info : | venkat_palleti.che@iipe.ac.in |
Details
Executive Summary : | With the increase of communication technology and Industry 4.0 the amount of data collected has increased from the industrial control systems such as oil, gas and chemical etc.. From these systems, sensor and actuator data is collected and use for future predictions to keep the system in a healthy state. Also with the increase of technology these systems are more vulnerable to attacks. Therefore, it is important to prevent the process from these attacks by detecting them in a timely manner in order to avoid serious damages to the process. The anomaly detectors will play a crucial role in keeping the process in safe and normal conditions. Machine learning/ artificial intelligence techniques will be applied on the time series of data to build anomaly detectors which will be used to predict the future values deviating from the normal scenario. The effectiveness of anomaly detectors will be assessed subject to the testing of different types of attacks. However, adversarial attacks which are designed to craft noise in the dataset samples may degrade the anomaly detector performance. Adversarial machine learning is basically the study of machine learning to defend anomaly detectors against potential attacks which might try to harm algorithms by confusing the model through increased noise in the dataset. These attacks tend to decrease the efficiency of the model by confusing them and pose danger to life and resources if employed in real life scenarios. Adversarial attacks could cause the most damage to hazardous industries like that of oil and gas, Chemicals, heavy machinery etc. Adversarial attacks have applications in image classification, audio signals, malware etc. However, less importance was given when it is applied to the industrial processes. In this work, we would like to use the standard benchmark Tennessee Eastman process dataset to study the impact of adversarial attacks on anomaly detectors of the process. We would like to make use of the anomaly detectors developed based on deep neural networks. The major contributions of the work are summarise as follows: 1. Propose threat models for adversarial attacks using white box gradient approach and construct adversarial data samples to deceive the anomaly detectors. 2. The adversarial attacks will be implemented on a standard Tennessee Eastman process dataset and test the efficiency of anomaly detectors. 3. Propose defense models to improve the robustness of the anomaly detectors against adversarial attacks. |
Total Budget (INR): | 6,60,000 |
Organizations involved