Date of Award

2026-08-01

Degree Name

Master of Science

Department

Computer Science

Advisor(s)

Aritran Piplai

Abstract

Machine learning (ML) models deployed in non-stationary environments must continually adapt to evolving data distributions. This challenge is particularly critical in cybersecurity, where malware, intrusion techniques, and adversarial behaviors evolve over time. Continual learning primarily enables incorporating new knowledge while preserving prior knowledge, however, indiscriminately retaining obsolete and harmful information can hinder future adaptation and consume limited model capacity. We argue that effective adaptation should not only acquire new knowledge, but also selectively discard obsolete and less useful historical knowledge before learning from a new distribution. In this work, we propose a meta-learning framework that learns what to forget to maximize future adaptation performance. At each distribution shift, a forgetting policy analyzes representations of historical training samples and predicts which samples should be unlearned before adaptation. The model is subsequently adapted to the new data, and its post-adaptation performance provides the reward used to optimize the forgetting policy across multiple episodes. Episodic optimization enables the policy to learn forgetting strategies that generalize to previously unseen distribution shifts. Experiments across multiple continual adaptation scenarios show that our approach consistently improves adaptation performance, providing an efficient alternative for resource- constrained ML systems.

Language

en

Provenance

Received from ProQuest

File Size

73 p.

File Format

application/pdf

Rights Holder

Daniel Lucio

Download

Share

COinS