Current Projects



Detection of private information in documents based on federated learning

The overall goal of this project is to train neural network models to extract private information from text including texts from scribed meetings, legal documents, medical data and so on.

Description

Data privacy is a becoming increasingly important with Data Protection Laws in Canada and the General Data Protection Regulation in European Union. The overall goal of this project is to train neural network models to extract private information from text including texts from scribed meetings, legal documents, medical data and so on. These models are supervised and require training data; however, training datasets are themselves subject to privacy. The outcome of this project will be a dataset of documents that contains private information of fictitious people and the labels that correspond to their private information as well as federated learning algorithms for detecting this private information.


Objectives

  • Software will be developed that will, for the application of interest, automatically collect pages from Internet and modify them to add privacy information. When private information is generated, the document will be annotated with the label that corresponds to a privacy type including name, email address, address, and affiliation and so on.
  • Development and training of the networks for federated learning that can be used to extract private information about people without exchanging data between the multiple sites.


Members

Rajitha Hathurusinghe rhath050@uottawa.ca, M.Sc. student
Dr. Isar Nehadgholi, NRC
Dr. Miodrag Bolic


Industrial Sponsor/Collaborator

IMRSV Data Labs and Mitacs
Funding: $40,000, 2018-2019


Drone detection and classification

In this project, we will develop a system for automated classification of UAVs and birds using artificial intelligence algorithms. The system will be based on processing data from radar sensor and sensor fusion of radar and camera data.

Description

Detection and tracking of UAVs using radars poses significant challenges because small UAVs typically have a small radar cross-section, fly at lower speed, and mostly at lower altitudes compared to larger aircraft. In addition, it is a nontrivial task to distinguish these small flying objects from birds. Signature detection from birds by itself is a difficult task, as data collection, distinguishing a single bird from a flock of flying birds and distinguishing different types of birds from a distance are all challenging tasks. Until now, several attempts to detect UAVs and distinguish them from birds using various sensors have achieved only limited success. In this project, we will develop a system for automated classification of UAVs and birds using artificial intelligence algorithms. The system will be based on processing data from radar sensor and sensor fusion of radar and camera data.


Members

Dr. Varun Mehta
Khurram Shafiq
Xu Zhang
Dr. Iraj Mantegh
Dr. Miodrag Bolic


Industrial Sponsor/Collaborator

NRC


Links and Papers

  • M. Bolic, Literature review on drone classification, 2018, unpublished.


AI Methods for Automated Software Testing

Clustering and prioritization can be done using modern artificial intelligence (AI) algorithms and this is exactly the objective of this project. Therefore, we are going to collect historical data after running test suite in real environment and then to use this data to train the AI models.

Description

In large companies, software codebase is very large and it is changing really fast. The teams spend a lot of time running tests even for very small changes in the code. To find the exact change that caused tests to fall, developers run every test at every change. However, this is very expensive and very time consuming.
Automated testing should help reducing testing time by prioritizing tests that have higher likelihood to detect problems and by clustering tests into groups so that it is not necessary to run all the tests in each group. Clustering and prioritization can be done using modern artificial intelligence (AI) algorithms and this is exactly the objective of this project. Therefore, we are going to collect historical data after running test suite in real environment and then to use this data to train the AI models. Our system will be able to perform online learning even after being deployed.
This project has several research challenges including development of AI algorithms, dealing with very large and complex software developed by Ericsson, and dealing with multiprocessing embedded system that runs on multiple boards.
The system should improve productivity of software developers in Ericsson significantly. In addition, it has potential to improve productivity of any software developers in Canada saving Canadian companies large amount of money.


Members

Dr. Varun Mehta
Dr. Miodrag Bolic


Industrial Sponsor/Collaborator

Ericsson
Funding: NSERC Engage Grant, $25,000, 2019-2020


Handling uncertainty in IOT systems

In this project we will address decision making with uncertainty in the cloud based on data coming from multiple sensors where sensors can represent Internet of Things (IoT) nodes, wearable medical devices or any other intelligent sensors on the edge.

Description

In this project we will address decision making with uncertainty in the cloud based on data coming from multiple sensors where sensors can represent Internet of Things (IoT) nodes, wearable medical devices or any other intelligent sensors on the edge. We are interested in providing decision with confidence levels in a given deadline and therefore our algorithms need to be scalable and computationally efficient.
Statistical data analytics and processing of sensor data is commonly based on linear and Gaussian mathematical models. These assumptions allow us to propagate uncertainty or to perform estimation of unknown parameters analytically. In case variables are statistically dependent, model is non-linear or non-Gaussian, it might not be possible to have analytical solution to the problem. With the advancement of computational statistics and Bayesian approach, processing data related to complex, non-linear and non-Gaussian systems has become feasible. Propagation of uncertainties based on Monte Carlo methods or other approximation methods can be universally applied and it is conceptually simple but can increase computational requirements by several orders of magnitude. Current solutions for implementing that Bayesian solutions are based on probabilistic programming. In this project, we will define a framework for probabilistic computing of sensor IoT data on the cloud. It will be based on Java and probabilistic programming language Figaro.
Several research problems are/will be tackled:

  1. Heterogeneous data from the devices will be collected and fused on the cloud. We will define Bayesian networks that would allow for combining heterogeneous data.
  2. Current published approaches for uncertainty quantification of large number of sensor data are based on simplified examples. We will propose scalable big data architecture for IoT systems. We will explore scalability of the approach presented in the research direction 1 in processing data on the cloud while optimizing usage of computational resource and achieving deadline for decision making.


Members

Mira Vrbaski
Dr. Shikharesh Majumdar
Dr. Miodrag Bolic


Industrial Sponsor/Collaborator

Funding: None


Published Papers and Reports

  • M. Vrbaski, M. Bolic, S. Majumdar, “Complex Event Recognition Notification Methodology for Uncertain IOT Systems Based on Micro-Service Architecture,” FiCloud 2018.