Cybersecurity & Real-Time Anomaly Detection:

We conduct Data-driven Cybersecurity research based on Machine Learning and Statistical Signal Processing. Our current work focuses on Real-Time Anomaly Detection, IoT Security, and Smart Grid Security. Specifically, we investigate quickly and accurately detecting and mitigating cyberattacks in challenging scenarios, e.g., high-dimensional, heterogeneous and dynamic systems, such as IoT networks, Smart Grid, and Intelligent Transportation Systems.

Students: Keval Doshi, Mahsa Mozaffari, Ammar Haydari, M. Necip Kurt (Columbia, EE), Elizabeth Hou (Michigan, ECE)
Sponsor: Southeastern Center for Electrical Engineering Education (SCEEE)
Representative papers:

  • M. Kurt, Y. Yilmaz and X. Wang, "Real-Time Detection of Hybrid and Stealthy Cyber-Attacks in Smart Grid", IEEE Transactions on Information Forensics and Security, Feb. 2019 [pdf]
  • E. Hou, Y. Yilmaz and A. Hero, "Anomaly Detection in Partially Observed Traffic Networks", IEEE Transactions on Signal Processing, 2019 [pdf]
  • M. Kurt, Y. Yilmaz and X. Wang, "Distributed Quickest Detection of Cyber-Attacks in Smart Grid", IEEE Transactions on Information Forensics and Security, Aug. 2018 [pdf]
  • A. Haydari and Y. Yilmaz, "Real-Time Detection and Mitigation of DDoS Attacks in Intelligent Transportation Systems", IEEE International Conference on Intelligent Transportation Systems (ITSC), 2018 [pdf]
  • Y. Yilmaz, "Online Nonparametric Anomaly Detection based on Geometric Entropy Minimization", IEEE International Symposium on Information Theory (ISIT), 2017 [pdf]
  • Y. Yilmaz and S. Uludag, "Mitigating IoT-based Cyberattacks on the Smart Grid", IEEE International Conference on Machine Learning and Applications (ICMLA), 2017 [pdf]

Resource Allocation & Reinforcement Learning:

Based on Markov decision processes (MDP) we derive optimum sequential decision making algorithms, e.g., for sequential estimation in high-dimensional settings, sequential joint detection and estimation, and agent-based scenario planning. We also derive practical solutions to MDP using reinforcement learning (RL) techniques for challenging high-dimensional and dynamic resource allocation problems, such as scenario planning for sea level rise, transformer aging and failure due to electric vehicle penetration in smart grid, resource allocation in fog radio access networks (Fog-RAN) for 5G IoT communications, and traffic light control in intelligent transportation systems.

The RL-based resource allocation solution in Fog-RAN adapts to the dynamic IoT environment as demonstrated by the figure below. The drops in the performance corresponds to the changes in the environment. The proposed RL algorithm quickly learns the optimum solution in the new environment.
Students: Almuthanna Nassar, Salman Sadiq Shuvo, Keval Doshi, Ammar Haydari
Sponsor: NSF

Project Website for Sea Level Rise Scenario Planning

Representative papers:
  • A. Nassar and Y. Yilmaz, "Resource Allocation in Fog RAN for Heterogeneous IoT Environments based on Reinforcement Learning", IEEE International Conference on Communications (ICC), 2019 [pdf]
  • Y. Yilmaz, S. Li and X. Wang, "Sequential joint detection and estimation: Optimum tests and applications," IEEE Transactions on Signal Processing, Oct. 2016. [pdf]
  • Y. Yilmaz, G. V. Moustakides and X. Wang, "Sequential and decentralized estimation of linear regression parameters in wireless sensor networks," IEEE Transactions on Aerospace and Electronic Systems, Feb. 2016. [pdf]

Multimodal Data Fusion:

We are interested in learning, in an unsupervised fashion, a low-dimensional feature vector for the observed instances from high-dimensional and heterogeneous data. In big data applications, it is common to have disparate data types, such as real-valued, integer-valued and categorical, together in the same dataset. Using generative Bayesian models we propose algorithms for fusing disparate data modalities from the exponential family of distributions in various applications, such as gene networks, social networks, recommender systems, and nuclear fuel cycles.

Our joint variational learning algorithm is capable of extracting shared communities across all graph layers as well as identifying communities unique to each layer. This figure shows for each chromosome the number of shared clusters, the number of private clusters for RNA gene expression levels and the number of private clusters for HiC contact maps between genes. This analysis suggests that the genes in chromosomes 2, 5, 15, 16, 20, 21, 22 are strongly co-expressed (high connectivity in RNA graph) and strongly connected (high connectivity HiC graph) since the number of shared clusters is dominant compared to the number of private clusters, while the genes in the other chromosomes are either strongly co-expressed or strongly connected.

Students: Mehmet Aktukmak, Hafiz Tiomoko Ali (CentraleSupelec, France)
Representative papers:

  • H. Ali, S. Liu, Y. Yilmaz, A. Hero, R. Couillet and I. Rajapakse, "Latent Heterogeneous Multilayer Community Detection", IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2019 [pdf]
  • M. Aktukmak, Y. Yilmaz and I. Uysal, "Matrix Factorization with Multimodal Side Information”, Neurocomputing, 2019
  • Y. Yilmaz and A. Hero, "Multimodal Event Detection in Twitter Hashtag Networks", Journal of Signal Processing Systems, 2018 [pdf]
  • Y. Yilmaz and A. Hero, "Multimodal Factor Analysis", IEEE International Workshop on Machine Learning for Signal Processing, 2015 [pdf]