A Comprehensive Framework for Targeted Transfer Learning: Privacy-Preserving, Secure, and Efficient Domain Adaptation
Year: 2026
Type: dissertation
Abstract: Machine learning models deployed in tactical environments rely on high-quality labeled datasets, which are often challenging, expensive, and time-consuming to obtain. In addition, formal techniques for assuring model performance under dynamic conditions remain limited. Transfer learning has emerged as a promising solution, but its effectiveness is limited by domain coverage and similarity between source and target datasets. This dissertation presents a comprehensive, modelagnostic Targeted Transfer Learning (TTL) framework that addresses transferability challenges between domains with minimal data, demonstrating how targeted sample selection improves adaptation across image classification, network security, and privacy-preserving collaborative settings. First, we develop the TTL framework for scenarios where direct access to datasets is restricted. This metadata-driven approach applies to a combinatorial coverage method to estimate domain correlation and guide sample selection. Evaluation on a satellite imagery dataset showed that targeted selection outperformed zero-shot transfer and random selection while conserving computational resources. Second, we investigate the TTL framework in settings where both datasets are directly available. This data-driven approach uses optimal transport distance to quantify distributional differences and guide selection of informative target samples. Validation across a broad set of benchmark image classification datasets demonstrated consistent improvements over conventional transfer learning. Third, we address network malware classification through Selective Targeted Transfer Learning (STTL), projecting source and target traffic data into a shared latent feature space and selecting relevant samples using a Kullback–Leibler divergence-based similarity measure. Experimental results across multiple attack-family scenarios showed that STTL achieved higher performance than conventional transfer learning, confirming that targeted selection generalizes across data types and domains. Finally, we extend the TTL framework through two decentralized privacy-preserving mechanisms. In a federated learning setting, clients share only summary statistics, and a server computes aggregation weights that enable knowledge transfer without exposing local data. In a blockchain-based setting, cryptographic verification secures model weights throughout the transfer process, while a smart-contract mechanism rewards high-quality participation. Both mechanisms consistently outperformed non-targeted baselines while preserving data confidentiality and integrity. This research establishes targeted transfer learning as a practical framework that enables machine learning models to adapt reliably, securely, and efficiently across image classification, network security, and distributed collaborative environments.
DOI: 10.25777/e3fd-gd70
Author: Ahamed, Sayyed Farid
Institution: Old Dominion University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225739907