New techniques and applications in deep-learning and machine learning for security
Dawn Song, Professor
Electrical Engineering and Computer Science
Applications for Fall 2024 are closed for this project.
Deep-learning has revolutionized AI, and achieves human-level accuracies in many tasks such as image recognition. Development on the optimization techniques and the availability of large amount of data make training a large architecture possible, which also opens up many intriguing problems to answer. In this project, we are interested in these deep learning problems in three general directions: (1) the mathematical property of deep learning models. For example, one important phenomenon discovered recently is that an attacker can manipulate very slightly the inputs to a deep learning classifier to make the prediction entirely wrong. We can study the robustness issue of deep learning systems against such adversarial manipulations. (2) applying deep learning techniques to novel application domains. For example, we will study program synthesis problems and security applications using deep learning. (3) The infrastructure to support large-scale deep learning. For example, we will study several existing efforts such as TensorFlow to design and develop new frameworks to make it easier for developers to build deep learning systems, architecture and applications. We are also open to other important and not well-understood deep learning related topics.
Moreover, machine learning techniques can be valuable in addressing security problems, e.g., identifying anomalous changes in behavior, clustering and classifying different behaviors/samples. We study how to apply machine learning techniques for a number of security applications including malware analysis and defense, social-network security analysis and defense.
Qualifications: solid background in math (probabilities) and machine learning,
proficient in programming in Python and C++
Hours: 12 or more hours
Engineering, Design & Technologies