From Training to Edge Inference: A Transfer Learning Pipeline for Low-Latency Malware Classification on Snapdragon-Enabled Samsung Galaxy S23 Ultra (Android 13)

Author (s)

Onyedika Christopher Agada

Abstract

There is an urgent need for a lightweight detection system that is able to operate efficiently without dependency on the cloud, because of the large number of edge devices in interconnected systems. This study introduces a framework that deploys a deep learning-based malicious software classifier on resource-constrained edge devices, utilizing malicious software binary files that are visualized as images and fine-tuning a ResNet18 architecture to recognize visual patterns. The system was further optimized using the Qualcomm AI Hub for deployment on a Snapdragon-powered Samsung Galaxy S23 Ultra (Android 13). A test accuracy of 97.47% was achieved, with 0.978 precision, 0.978 recall, and an F1-score of 0.978. Additionally, a 75% model size reduction (44.8 – 11.2 MB) was achieved, and an on-device inference latency of 1 millisecond (batch size = 1) which is excellent for real-time malware detection. This research advances real-time threat mitigation in edge devices and enables scalable cyber defenses that preserve privacy.

Key words: Deep Learning, Edge computing, Image Classification, Machine learning, Malware Detection, Transfer Learning.

Download PDF

Cite This Article:

Agada, O.C. (2025). From Training to Edge Inference: A Transfer Learning Pipeline for Low-Latency Malware Classification on Snapdragon-Enabled Samsung Galaxy S23 Ultra (Android 13), Journal of Scientific Reports, 11(1), 51-64. DOI: https://doi.org/10.58970/JSR.1135

About Author (s)

Onyedika Christopher Agada, Department of Cybersecurity, Uskudar University, Istanbul, Turkey. ORCID: https://orcid.org/0009-0003-4337-597X

Download PDF

DOI: https://doi.org/10.58970/JSR.1135

Summarize this article with: