Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Fabio Di Troia
Second Advisor
William Andreopoulos
Third Advisor
Vuthea Chheang
Keywords
Federated Learning, Homomorphic Encryption, Malware Classification, Public Key Infrastructure, Convolutional Neural Networks, Zero-Trust Architecture
Abstract
Federated Learning (FL) is an algorithmic approach that allows the distributed training of sensitive machine learning models, such as malware classifiers, without revealing the underlying datasets. FL achieves this by transmitting local model param eters (weights) rather than raw data. Unfortunately, one of the most commonly used federated approaches today, Differential Privacy (DP), introduces severe limitations. DP enforces privacy by injecting statistical noise into high-dimensional gradients, which severely degrades gradient quality and destroys the subtle features needed for accurate classification. This project introduces a secure and reliable FL framework using Homomorphic Encryption (HE) to preserve model utility. Experimental evaluation shows that this framework achieves 97.76 percent classification accuracy, matching a centrally managed, unencrypted plaintext baseline while providing deterministic protection from Sybil and poisoning attacks. The "Fail-Closed" security posture for the system ensures cryptographically correct data, effectively confining potential breaches of either confidentiality or integrity to non-catastrophic, localized Denial of Service (DoS) events.
Recommended Citation
Challa, Chandrakanth Reddy, "Secure and Resilient Federated Learning for Malware Classification" (2026). Master's Projects. 1818.
DOI: https://doi.org/10.31979/etd.xzp5-4wv4
https://scholarworks.sjsu.edu/etd_projects/1818
