UMBC Cybersecurity and Artificial Intelligence - UMB ICTR Core
UMBC is proud to become a part of UMB’s ICTR
UMBC will provide a new Cybersecurity and Artificial Intelligence Core to the UMB ICTR community as part of the UMB ICTR Informatics Core. In turn, UMBC faculty will become eligible for many of the existing ICTR core offerings.
For more information on the UMBC-ICTR Core, please email ICTR-Navigator@umaryland.edu.
To apply for services through the UMB ICTR, go to Investigator Resources
UMBC-ICTR Core Capabilities:
- Securing medical devices
- Securing smart systems, e.g., smart surgery rooms
- Protecting learned/predictive models from attacks
- Deep learning and artificial neural networks
- Natural language processing
- Graph analytics
- Time series analytics
- Data visualization augmented reality and virtual reality
UMBC-ICTR Example Services:
- Consult to uncover possible cybersecurity risks associated with devices and/or systems.
- Consult on ways to protect devices and/or systems from attackers, either at design time or after deployment.
- Consult on ways to apply artificial intelligence and machine learning to solve specific problems given existing data, including processing pipelines, specific algorithms, and evaluation methodology
- Advise on what additional data could be collected or obtained to potentially improve the utility of AI/ML for specific use cases.
- Construct simple proof-of-concept AI/ML systems to understand what level of performance might be achieved with more time, data, or resources
- Advise on the best visual representations to explore complex datasets and to communicate results to others.
The University of Maryland, Baltimore County brings expertise in two main areas to the UMB ICTR: Cybersecurity and Artificial Intelligence/Machine Learning. These two areas are often interdependent, with data-driven methods applied to problems in cybersecurity, and work on securing AI and ML systems against data-oriented attacks. Faculty at UMBC also have significant experience in application domains relevant to the UMB ICTR, such as medical imaging, diagnosis from temporal data such as EEG and vital signs, and embedded device security. Below is a list of topics within the two broad areas, with descriptions of UMBC capabilities and, where applicable, example applications. We start with broad overviews of capabilities in cybersecurity and AI/ML, and the more focused sub-topics are by no means an exhaustive list.
UMBC has broad capabilities in the general area of cybersecurity. These range from security of chips, devices and embedded systems to network security to software security and all the way up to security of systems and policies and procedures around them. The UMBCCenter for Cybersecurity is an interdisciplinary university center that unifies UMBC’s those capabilities. The Center aims to provide both Maryland and the nation with academic and research leadership, collaboration, innovation, and outreach in this critical discipline by streamlining our academic, research, workforce development, and technology incubation activities. UMBC embraces a holistic approach to cybersecurity as a key component of National Security and views it as more than just defending computers and networks. The UMBC Center for Cybersecurity draws upon UMBC’s expansive range of academic and research faculty expertise in not just the computing disciplines (computer science, computer engineering, electrical engineering, information systems), but also in the natural, mathematical and social sciences to explore the many technical, social, policy, and business challenges facing the cybersecurity profession.
The medical field generates data, whether it comes from clinical trials or radiology or scientific papers or patient notes. The data come in different modalities (numeric, time series, 2D images, 3D volumes, text) and are often inter-related, e.g., a patient may generate a vital signs time series and a CT scan, with notes written about the patient’s condition and what the data say about that condition. UMBC has long-standing expertise in AI and ML to support reasoning and decision-making over many different time ranges. For example, we’ve built systems for diagnosis using radiology images, identifying operating room state from video to assist with OR scheduling, and predicting hospital readmission from vital signs data.
Many embedded devices, such as infusion pumps are made of parts from a variety of vendors. Some are even implantable. A significant security risk is trojan hardware or software running on those parts, perhaps installed at the factory or elsewhere in the supply chain, that can lead to information leakage or malicious behavior. Some of these devices are wirelessly connected, and that channel can be used to subvert them. UMBC faculty have developed novel methods for non-invasive measurement of device behavior (e.g., by looking at power draw over very small time scales) to identify the presence of unwanted hardware/software modifications on devices. The primary focus is on low-power and embedded devices.
An increasing number of medical systems are CyberPhysical, i.e., they have sensors ingesting data, computing, and then controlling physical systems. Smart surgery rooms are a good example of such systems. Attackers can cause significant physical damage, such as changing the outcome of a surgery, by attacking the system. For example, detecting the cystic artery within the Calot’s triangle is an important step in Cholecystectomy. One can attack the image detection algorithm of a smart surgery system to interfere in this step and cause problems. UMBC Faculty have developed novel methods to detect and mitigate attacks on such systems.
As data-driven, learned models play larger roles in areas like self-driving cars and medical diagnosis, the security of those models becomes vitally important. However, many such models have recently been discovered to be susceptible to adversarial attacks, such as visually imperceptible changes to a few pixels causing a classifier to consistently make mistakes(e.g., making a radiology image with a clear tumor be classified as healthy). UMBC faculty have expertise in both generating and protecting against such attacks on learned models based on a variety of modalities.
A particular strength of UMBC is in the area of deep learning, which is revolutionizing how organizations extract value from large, complex datasets. Our work spans the range of theoretical to practical, with application domains as diverse as object tracking and activity recognition in video, generating textual descriptions of images and videos, exploring the relationship between functional brain activities during moral decision making and psychopathic personality traits, and enabling deep learning on low-power embedded devices.
A tremendous amount of the world’s information is stored as text, including medical records, scientific articles, legal opinions, and log files from various computer systems. UMBC faculty have deep experience in information and relation extraction to populate knowledge-bases or knowledge graphs with facts found automatically in unstructured text; text summarization; extracting stereotypical event sequences from text; joint modeling of text and other modalities like images; and text classification. For example, UMBC Faculty have worked with the VA to read a vast amount of patient care notes regarding, and automatically identify key information (pain variation over time, pain related to medication, etc)and summarize these for the attending physician.
Graphs are the ideal representation for many domains, including protein signaling networks, the interactions of entities on computer networks, and flows of patients through the various touch points involved in getting treatment in hospitals. UMBC faculty have expertise in a variety of types of graph analytics, including models of sets of graphs and distributions over graphs for graph classification, making flexible judgments of similarity between graphs using deep neural networks, applications of graph analytics to network security, and ontology-backed reasoning over large knowledge graphs.
Blockchain technology is finding applications in a wide variety of domains beyond cryptocurrency, including finance, supply chain, manufacturing, and healthcare. UMBC researchers are developing both new blockchain paradigms (e.g., permissioned blockchains)and exploring novel applications of blockchain technology.
Temporal data is abundant, including EEG, speech, vital signs, values of financial instruments, supply levels, etc. UMBC faculty have expertise in time series visualization, classification, pattern and anomaly discovery, and prediction.
UMBC has deep expertise in graphics and visualization, making sense of large volumes of data by representing them visually, including in Immersive/Cave like environments or using gamification. Work on HCC atUMBC is driven by a diverse collection of interrelated research questions centered on the design, implementation, and evaluation of highly-usable interactive systems. Data visualization is central to this work. We employ a variety of methodologies including both quantitative and qualitative approaches, lab and field-based data collection, and usability engineering approaches including user-centered design, participatory design, and other related techniques. The three core areas are: Accessible Computing, broadly defined to include issues associated with disabilities, age, culture, and context-aware computing; Human-Information Interaction, which studies information behavior and the design of user interaction methods to support that behavior; and Social Computing, which studies social behavior as it relates to computational systems and evaluating the various environments therein.