Machine learning has opened up numerous opportunities and applications for vari-ous networking applications. This dissertation focuses on ML-based solutions in constrained
environments for different networking applications.
First, we present a framework that detects and summarizes key global eventsfrom distributed crowd-sensed data in a bandwidth-constrained environment. We introduce
BigEye, a novel framework that only transfers limited data from distributed producers to a
central summarizer, supporting highly accurate detection and concise visual summarization
of key events of global interest.
Second, we develop AcTrak, a framework to control steerable cameras through anetwork to retrieve telemetrics of interest. AcTrak automates a camera’s motion to switch
appropriately between zooming in on existing targets in a scene to track their activities and
zooming out to search for new targets arriving in the area of interest. We aim to achieve
a good trade-off between these two tasks, ensuring that new targets are observed by the
camera before they leave the scene while frequently monitoring the activities of existing
targets.
Third, we uncover a vulnerability that enables fast and stealthy data exfiltrationover DNS channels. While existing defenses against such attacks appear robust, we demon-
strate that our carefully designed and novel DNS exfiltration attack, Dolos, which uses
a generative adversarial network (GAN), can encode sensitive data to evade these detec-
tors unlike existing state-of-the-art attack methods. Additionally, Dolos can significantly
expedite exfiltration compared to prior methods.
