Multi-Task Learning for Federated Security Applications
Learning diverse tasks in device type-specific anomaly detection models
This project is only available to UTSA PhD students.
Systems based on distributed machine learning like federated learning have been proposed for various security applications. Federated learning which aggregates models from different participating clients is particularly advantageous for applications like intrusion detection for IoT networks, since it allows to aggregate device type-specific models for anomaly detection even for devices that typically generate only little data. One inherent problem with such set-ups is, however, that depending on the usage context of specific devices, their behaviour may exhibit markedly different benign patterns. For example, depending on whether an IP security camera with a motion detector is used to monitor the exterior of a house or a particular room inside, the usage pattern of the camera will be quite different. This means that the benign behaviour model of a single device type may consist of a number of different ‘tasks’ that the device exhibits in different contexts. For anomaly detection purposes it is therefore not sufficient to train a model that covers the most common task, but the model should accurately capture all behavioural tasks that are benign.
In this project, the task would be to investigate methods with which device type-specific models can be trained taking the multi-task nature of benign behaviour of monitored objects into account. The target system scenario would be IoT intrusion detection.
The project is supervised by Prof. Dr.-Ing. Ahmad-Reza Sadeghi.
Additional information
Capacity | open-ended |
Project available until | Spring 2023 |
Credits | 18 ECTS |
Available via Remote | Yes |