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

Aditya Pandey

Keywords

Malware Classification, Lightweight Neural Networks, SHAP, Adversarial Machine Learning, LSTM, Network Traffic Analysis

Abstract

This project mainly makes use of light weight neural networks and temporal deep learning models to investigate explainable and adversarial robust multi-class IoT malware classification. The usage of Internet of Things (IoT) has continued to expand in multiple areas like smart homes, health care, industrial systems and also into campus networks, but the hardware limitations of such devices make the deployment of many heavyweight defence mechanisms difficult in real-time. While the binary classification of the malware wherein we identify benign-versus-malware is useful for initial screening, but many operational security workflows need clear categorization of the attack type inorder to perform defensive actions that are tailored to threats. This work mainly focusses on extending a lightweight sparse neural network from a binary classification setting into a multi-class malware detection framework that is also more efficient and robust making use of the IoT-23 network traffic dataset. The implemented framework mainly consisted of 3 stages. First, the entire dataset is normalized, cleaned and then transformed into 6 classes of different malware types. Second, a lightweight Sparse Neural Network with Iterative Pruning (SNIPE) architecture is trained as the primary tabular classifier, then after the initial training SHAP- guided feature analysis has been performed in order to refine the feature subset and to determine the most influential traffic attributes. Third, the optimized SNIPE model based on SHAP is reimplemented in Keras in order to support ART framework for adversarial robustness evaluation and training, after which this model is compared with multiple temporal and sequence-based recurrent models constructed from the temporal windows of the source-IP flows. In addition to accuracy, the study mainly emphasizes on interpretability, robustness, compactness and practicality. On the clean multi-class classification, the baseline SNIPE has given an accuracy of 89.33%, while the optimized SNIPE, based on the SHAP analysis and reimplemented in Keras has achieved 95.46% accuracy on the clean test set and was used as the baseline for the evaluation of adversarial attacks using FGSM, PGD,C&W methods. A completely adversarially trained robust SNIPE model maintained an accuracy of 93.20% on the clean data. The robust model has substantially improved resistance to strongest attacks, including improved 49.40% accuracy under PGD with ε = 0.3. In order to model the temporal behaviour, source-Ip grouped flow windows were used to train the LSTM and RNN models. LSTM achieved stronger overall performance, reaching 98.78% clean accuracy and an accuracy of 64.90% under PGD ε = 0.3 after the adversarial training. A hybrid SNIPE-LSTM ensemble was also created and it achieved an unified-test-set performance of 98.98% accuracy. Major contribution of this projects are: (1) a practical approach for reformulation of multiclass malware detection in IoT devices , (2) a SHAP based optimization workflow that helped improve interpretation and model selection, (3) an adversarial evaluation and training workflow that helped improve the robustness, and (4) also a comparison between multiple tabular and temporal sequence models that shows the trade-off between the robustness and compactness. The resulting pipeline provided a strong baseline for future IoT systems that need to balance between inference times, interpretability and resilience against adversarial attacks.

Download

Available for download on Thursday, May 20, 2027

Share

COinS