Fabio Di Troia
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Fabio Di Troia has a total of 89 co-authors. Here are the top co-authors by number of shared publications:
- Mark Stamp
Department of Computer Science San Jose State University
23 shared publications Mark Stamp
13 shared publications- Martin Jureček
Czech Technical University in Prague
5 shared publications - Thomas H. Austin
San Jose State University
5 shared publications Corrado Aaron Visaggio
Department of Engineering Universita degli Studi del Sannio
3 shared publicationsKaterina Potika
3 shared publications
Latest Publications
- Open Access
- 2 November 2026
- Published by MDPI AG in Preprints.org
Revisiting Opcode Representations for Malware Family Classification: Learned Embeddings Versus Frequency Baselines
Opcode representations provide numerical features for malware-family classification, but their value must be assessed against simple distributional controls under a shared evaluation protocol. We compare globally trained Word2Vec, GloVe, and fastText embeddings with normalized opcode frequencies on 1,045 files from nine malware families. Ten stratified repeated holdouts use identical file assignments across representations. Embedding epochs are selected using an inner validation partition, followed by a fresh fit on the outer training files. A fixed Random Forest evaluates original-only training and conventional oversampling on the same held-out files. Mean original-only accuracies for opcode frequency, Word2Vec, GloVe, and fastText were 1.0000, 0.9995, 0.9981, 0.9976, respectively. The frequency baseline matched or exceeded the learned representations’ mean accuracy. SMOTE increased macro F1 in 4 of 40 representation–split pairs (none for frequency, which was already at the ceiling). All four representations achieved near-ceiling accuracy on this selected closed-set benchmark.
- 2 November 2026
- Published by MDPI AG in Preprints.org
Open AccessRevisiting Opcode Representations for Malware Family Classification: Learned Embeddings Versus Frequency Baselines
Opcode representations provide numerical features for malware-family classification, but their value must be assessed against simple distributional controls under a shared evaluation protocol. We compare globally trained Word2Vec, GloVe, and fastText embeddings with normalized opcode frequencies on 1,045 files from nine malware families. Ten stratified repeated holdouts use identical file assignments across representations. Embedding epochs are selected using an inner validation partition, followed by a fresh fit on the outer training files. A fixed Random Forest evaluates original-only training and conventional oversampling on the same held-out files. Mean original-only accuracies for opcode frequency, Word2Vec, GloVe, and fastText were 1.0000, 0.9995, 0.9981, 0.9976, respectively. The frequency baseline matched or exceeded the learned representations’ mean accuracy. SMOTE increased macro F1 in 4 of 40 representation–split pairs (none for frequency, which was already at the ceiling). All four representations achieved near-ceiling accuracy on this selected closed-set benchmark.
- 24 October 2026
- Published by Springer Nature in Journal of Computer Virology and Hacking Techniques
- Vol. 22 (1) , 99
Classifier-guided discrete diffusion for malware family-conditional graph generation
- 24 October 2026
- Published by Springer Nature in Journal of Computer Virology and Hacking Techniques
- Vol. 22 (1) , 99
Classifier-guided discrete diffusion for malware family-conditional graph generation
- Open Access
- 3 September 2026
- Published by MDPI AG in Electronics
- Vol. 15 (17) , 3969
Integrating Opcode N-Grams and Word Embeddings for Enhanced Malware Classification: A Comparative Study with Transformer-Based Representations
This work proposes a comparative framework for malware classification that evaluates the synergy between traditional feature engineering and modern deep learning architectures. Our methodology follows two primary paths: first, we integrate opcode n-grams with word-embedding techniques (Word2Vec, Doc2Vec, and FastText) to capture local execution patterns in dense vector spaces. Second, we evaluate end-to-end representations using transformer-based models (BERT and ViT) and a raw opcode-based 1D Convolutional Neural Network (1D-CNN) to determine if effective features can be learned without explicit n-gram engineering. Both pathways are rigorously tested across a suite of classifiers, including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbor (k-NN), and CNNs. Experimental results for multi-class classification demonstrate that while transformer-based models offer high automated feature extraction capabilities, the combination of opcode n-grams with word embeddings remains a highly effective and interpretable approach for detecting real-world malware.
- 3 September 2026
- Published by MDPI AG in Electronics
- Vol. 15 (17) , 3969
Open AccessIntegrating Opcode N-Grams and Word Embeddings for Enhanced Malware Classification: A Comparative Study with Transformer-Based Representations
This work proposes a comparative framework for malware classification that evaluates the synergy between traditional feature engineering and modern deep learning architectures. Our methodology follows two primary paths: first, we integrate opcode n-grams with word-embedding techniques (Word2Vec, Doc2Vec, and FastText) to capture local execution patterns in dense vector spaces. Second, we evaluate end-to-end representations using transformer-based models (BERT and ViT) and a raw opcode-based 1D Convolutional Neural Network (1D-CNN) to determine if effective features can be learned without explicit n-gram engineering. Both pathways are rigorously tested across a suite of classifiers, including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbor (k-NN), and CNNs. Experimental results for multi-class classification demonstrate that while transformer-based models offer high automated feature extraction capabilities, the combination of opcode n-grams with word embeddings remains a highly effective and interpretable approach for detecting real-world malware.
- Open Access
- 30 June 2026
- Published by MDPI AG in Electronics
- Vol. 15 (13) , 2848
Addressing Data Scarcity in Malware Classification via Pixel-Level Synthetic Image Generation
Deep learning-based malware classification using image representations has emerged as a highly effective paradigm for threat detection. However, training robust neural networks is frequently bottlenecked by data scarcity and severe class imbalances in real-world repositories. This study investigates the viability of using an autoregressive PixelCNN framework to synthesize high-fidelity, class-specific malware images to augment limited training distributions. Utilizing the benchmark Malimg dataset, we systematically evaluate a Convolutional Neural Network (CNN) classifier across varying ratios of synthetic-to-authentic data under strict data scarcity constraints (ranging from 10 to 80 authentic samples per family). Our experimental results reveal that while PixelCNN successfully replicates intricate, byte-level micro-textures, classifiers trained exclusively on synthetic data experience catastrophic performance degradation, yielding an accuracy of just 3%. Crucially, however, the introduction of a minimal authentic data anchor (15% to 20%) restores functional decision boundaries, immediately elevating classification accuracy up to 72%. Furthermore, performance saturates rapidly once the training matrix reaches a 50/50 synthetic-to-authentic split, achieving up to 82% classification accuracy, rendering it highly competitive with the 89% accuracy upper bound of a fully authentic baseline. These findings demonstrate an exceptional degree of data efficiency, proving that generative autoregressive augmentation can halve the authentic data collection burden in cybersecurity workflows provided a minor, real-world baseline anchor is preserved.
- 30 June 2026
- Published by MDPI AG in Electronics
- Vol. 15 (13) , 2848
Open AccessAddressing Data Scarcity in Malware Classification via Pixel-Level Synthetic Image Generation
Deep learning-based malware classification using image representations has emerged as a highly effective paradigm for threat detection. However, training robust neural networks is frequently bottlenecked by data scarcity and severe class imbalances in real-world repositories. This study investigates the viability of using an autoregressive PixelCNN framework to synthesize high-fidelity, class-specific malware images to augment limited training distributions. Utilizing the benchmark Malimg dataset, we systematically evaluate a Convolutional Neural Network (CNN) classifier across varying ratios of synthetic-to-authentic data under strict data scarcity constraints (ranging from 10 to 80 authentic samples per family). Our experimental results reveal that while PixelCNN successfully replicates intricate, byte-level micro-textures, classifiers trained exclusively on synthetic data experience catastrophic performance degradation, yielding an accuracy of just 3%. Crucially, however, the introduction of a minimal authentic data anchor (15% to 20%) restores functional decision boundaries, immediately elevating classification accuracy up to 72%. Furthermore, performance saturates rapidly once the training matrix reaches a 50/50 synthetic-to-authentic split, achieving up to 82% classification accuracy, rendering it highly competitive with the 89% accuracy upper bound of a fully authentic baseline. These findings demonstrate an exceptional degree of data efficiency, proving that generative autoregressive augmentation can halve the authentic data collection burden in cybersecurity workflows provided a minor, real-world baseline anchor is preserved.
- Open Access
- 4 December 2025
- Published by Springer Nature in Research Square
A Comparative Study of Linear and Non-Linear Dimensionality Reduction for Opcode-Frequency Malware Classification
High-dimensional feature spaces in malware classification pose significant challenges for machine learning performance. To address these challenges, this paper presents a comparative evaluation of four dimensionality-reduction techniques—Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Uniform Manifold Approximation and Projection (UMAP), and Autoencoder-based reduction—applied to opcode-frequency representations of malware. Using a corpus comprising 82,569 samples and 1,796 opcodes, we analyze the effect of each reduction method across multiple target dimensions and two classifier architectures: Extreme Gradient Boosting (XGBoost) and a three-layer Multilayer Perceptron (MLP). Results show that LDA achieves strong separability at lower dimensions, while PCA performs best at higher dimensions where variance preservation is critical. Autoencoder-based reduction provides consistently high accuracy with compact representations, whereas UMAP exhibits limited benefits for tabular opcode data. The findings highlight trade-offs between linear and non-linear reduction strategies and provide guidance for selecting efficient feature compression methods in large-scale malware analysis.
- 4 December 2025
- Published by Springer Nature in Research Square
Open AccessA Comparative Study of Linear and Non-Linear Dimensionality Reduction for Opcode-Frequency Malware Classification
High-dimensional feature spaces in malware classification pose significant challenges for machine learning performance. To address these challenges, this paper presents a comparative evaluation of four dimensionality-reduction techniques—Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Uniform Manifold Approximation and Projection (UMAP), and Autoencoder-based reduction—applied to opcode-frequency representations of malware. Using a corpus comprising 82,569 samples and 1,796 opcodes, we analyze the effect of each reduction method across multiple target dimensions and two classifier architectures: Extreme Gradient Boosting (XGBoost) and a three-layer Multilayer Perceptron (MLP). Results show that LDA achieves strong separability at lower dimensions, while PCA performs best at higher dimensions where variance preservation is critical. Autoencoder-based reduction provides consistently high accuracy with compact representations, whereas UMAP exhibits limited benefits for tabular opcode data. The findings highlight trade-offs between linear and non-linear reduction strategies and provide guidance for selecting efficient feature compression methods in large-scale malware analysis.
- Open Access
- 5 November 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (21) , 11791
Embedding-Driven Synthetic Malware Generation with Autoencoders and Cluster-Tangent Diffusion
Malware has become increasingly sophisticated over the years, with zero-day attacks emerging at an alarming pace. Effective detection and analysis demand real malware samples, which are expensive and skill-dependent to extract. As a result, generating high quality synthetic samples from scarce data sets becomes a crucial method for strengthening detection software. This paper focuses on presenting generation techniques that optimize the embedding space to produce high-quality synthetic samples, even under constrained datasets. The dataset used in this paper consists of 500 Windows malware API call samples that were processed using embedding and Generative AI (Gen AI) techniques to generate synthetic malware. Two novel contributions are highlighted in this paper. (1) The integration of autoencoders with pretrained NLP models (BERT and ELMo) to enhance the quality of embeddings. Autoencoders extract features and learn patterns from the data to generate higher-quality embeddings than those generated using other techniques alone. (2) Cluster-Tangent Diffusion (CT-Diff): a novel application of manifold diffusion. Manifold diffusion improves upon diffusion and other Gen AI techniques by focusing on generating samples along the distribution of the original data using structured noise instead of standard gaussian noise. Collectively these two contributions have consistently outperformed previous techniques. Furthermore, the results demonstrate the feasibility of generating reliable fake samples even in low data scenarios.
- 5 November 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (21) , 11791
Open AccessEmbedding-Driven Synthetic Malware Generation with Autoencoders and Cluster-Tangent Diffusion
Malware has become increasingly sophisticated over the years, with zero-day attacks emerging at an alarming pace. Effective detection and analysis demand real malware samples, which are expensive and skill-dependent to extract. As a result, generating high quality synthetic samples from scarce data sets becomes a crucial method for strengthening detection software. This paper focuses on presenting generation techniques that optimize the embedding space to produce high-quality synthetic samples, even under constrained datasets. The dataset used in this paper consists of 500 Windows malware API call samples that were processed using embedding and Generative AI (Gen AI) techniques to generate synthetic malware. Two novel contributions are highlighted in this paper. (1) The integration of autoencoders with pretrained NLP models (BERT and ELMo) to enhance the quality of embeddings. Autoencoders extract features and learn patterns from the data to generate higher-quality embeddings than those generated using other techniques alone. (2) Cluster-Tangent Diffusion (CT-Diff): a novel application of manifold diffusion. Manifold diffusion improves upon diffusion and other Gen AI techniques by focusing on generating samples along the distribution of the original data using structured noise instead of standard gaussian noise. Collectively these two contributions have consistently outperformed previous techniques. Furthermore, the results demonstrate the feasibility of generating reliable fake samples even in low data scenarios.
- Open Access
- 4 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (19) , 3937
Discriminative Regions and Adversarial Sensitivity in CNN-Based Malware Image Classification
The escalating prevalence of malware poses a significant threat to digital infrastructure, demanding robust yet efficient detection methods. In this study, we evaluate multiple Convolutional Neural Network (CNN) architectures, including basic CNN, LeNet, AlexNet, GoogLeNet, and DenseNet, on a dataset of 11,000 malware images spanning 452 families. Our experiments demonstrate that CNN models can achieve reliable classification performance across both multiclass and binary tasks. However, we also uncover a critical weakness in that even minimal image perturbations, such as pixel modification lower than 1% of the total image pixels, drastically degrade accuracy and reveal CNNs’ fragility in adversarial settings. A key contribution of this work is spatial analysis of malware images, revealing that discriminative features concentrate disproportionately in the bottom-left quadrant. This spatial bias likely reflects semantic structure, as malware payload information often resides near the end of binary files when rasterized. Notably, models trained in this region outperform those trained in other sections, underscoring the importance of spatial awareness in malware classification. Taken together, our results reveal that CNN-based malware classifiers are simultaneously effective and vulnerable to learning strong representations but sensitive to both subtle perturbations and positional bias. These findings highlight the need for future detection systems that integrate robustness to noise with resilience against spatial distortions to ensure reliability in real-world adversarial environments.
- 4 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (19) , 3937
Open AccessDiscriminative Regions and Adversarial Sensitivity in CNN-Based Malware Image Classification
The escalating prevalence of malware poses a significant threat to digital infrastructure, demanding robust yet efficient detection methods. In this study, we evaluate multiple Convolutional Neural Network (CNN) architectures, including basic CNN, LeNet, AlexNet, GoogLeNet, and DenseNet, on a dataset of 11,000 malware images spanning 452 families. Our experiments demonstrate that CNN models can achieve reliable classification performance across both multiclass and binary tasks. However, we also uncover a critical weakness in that even minimal image perturbations, such as pixel modification lower than 1% of the total image pixels, drastically degrade accuracy and reveal CNNs’ fragility in adversarial settings. A key contribution of this work is spatial analysis of malware images, revealing that discriminative features concentrate disproportionately in the bottom-left quadrant. This spatial bias likely reflects semantic structure, as malware payload information often resides near the end of binary files when rasterized. Notably, models trained in this region outperform those trained in other sections, underscoring the importance of spatial awareness in malware classification. Taken together, our results reveal that CNN-based malware classifiers are simultaneously effective and vulnerable to learning strong representations but sensitive to both subtle perturbations and positional bias. These findings highlight the need for future detection systems that integrate robustness to noise with resilience against spatial distortions to ensure reliability in real-world adversarial environments.
- Open Access
- 1 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (19) , 3915
Robust Hashing for Improved CNN Performance in Image-Based Malware Detection
This paper presents a comparative study on the impact of robust hashing in enhancing image-based malware classification. While Convolutional Neural Networks (CNNs) have shown promise when working with image-based malware samples, their performance degrades significantly when obfuscation techniques are taken into consideration to hamper the malware classification or detection. To address this, we apply a robust hashing technique that generates invariant visual representations of malware samples, enabling improved generalization under obfuscation implemented as image salting. Using a custom obfuscation method to simulate polymorphic variants, we evaluate MobileNet, ResNet, and DenseNet architectures across five salting conditions (0% to 40%). The results demonstrate that robust hashing substantially boosts classification accuracy, with DenseNet achieving 89.50% on unsalted data, compared to only 68.00% without hashing. Across all salting levels, models consistently performed better when robust hashing was applied, confirming its effectiveness in preserving structural features and mitigating adversarial noise. These findings position robust hashing as a powerful preprocessing strategy for resilient malware detection.
- 1 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (19) , 3915
Open AccessRobust Hashing for Improved CNN Performance in Image-Based Malware Detection
This paper presents a comparative study on the impact of robust hashing in enhancing image-based malware classification. While Convolutional Neural Networks (CNNs) have shown promise when working with image-based malware samples, their performance degrades significantly when obfuscation techniques are taken into consideration to hamper the malware classification or detection. To address this, we apply a robust hashing technique that generates invariant visual representations of malware samples, enabling improved generalization under obfuscation implemented as image salting. Using a custom obfuscation method to simulate polymorphic variants, we evaluate MobileNet, ResNet, and DenseNet architectures across five salting conditions (0% to 40%). The results demonstrate that robust hashing substantially boosts classification accuracy, with DenseNet achieving 89.50% on unsalted data, compared to only 68.00% without hashing. Across all salting levels, models consistently performed better when robust hashing was applied, confirming its effectiveness in preserving structural features and mitigating adversarial noise. These findings position robust hashing as a powerful preprocessing strategy for resilient malware detection.
- Open Access
- 20 August 2025
- Published by MDPI AG in Preprints.org
Robust Hashing for Improved CNN Performance in Image-Based Malware Detection
This paper presents a comparative study on the impact of robust hashing in enhancing image-based malware classification. While convolutional neural networks (CNNs) have shown promise when working with image-based malware samples, their performance degrades significantly when obfuscation techniques are taken in consideration to hamper the malware classification or detection. To address this, we apply a robust hashing technique that generates invariant visual representations of malware samples, enabling improved generalization under obfuscation implemented as image salting. Using a custom obfuscation method to simulate polymorphic variants, we evaluate MobileNet, ResNet, and DenseNet architectures across five salting conditions (0% to 40%). The results demonstrate that robust hashing substantially boosts classification accuracy, with DenseNet achieving 89.50% on unsalted data, compared to only 68.00% without hashing. Across all salting levels, models consistently performed better when robust hashing was applied, confirming its effectiveness in preserving structural features and mitigating adversarial noise. These findings position robust hashing as a powerful preprocessing strategy for resilient malware detection. Keywords: Malware Detection, Convolutional Neural Networks, Robust Hashing, Dataset Obfuscation, MobileNet, ResNet, DenseNet, Adversarial Attacks, Cybersecurity.
- 20 August 2025
- Published by MDPI AG in Preprints.org
Open AccessRobust Hashing for Improved CNN Performance in Image-Based Malware Detection
This paper presents a comparative study on the impact of robust hashing in enhancing image-based malware classification. While convolutional neural networks (CNNs) have shown promise when working with image-based malware samples, their performance degrades significantly when obfuscation techniques are taken in consideration to hamper the malware classification or detection. To address this, we apply a robust hashing technique that generates invariant visual representations of malware samples, enabling improved generalization under obfuscation implemented as image salting. Using a custom obfuscation method to simulate polymorphic variants, we evaluate MobileNet, ResNet, and DenseNet architectures across five salting conditions (0% to 40%). The results demonstrate that robust hashing substantially boosts classification accuracy, with DenseNet achieving 89.50% on unsalted data, compared to only 68.00% without hashing. Across all salting levels, models consistently performed better when robust hashing was applied, confirming its effectiveness in preserving structural features and mitigating adversarial noise. These findings position robust hashing as a powerful preprocessing strategy for resilient malware detection. Keywords: Malware Detection, Convolutional Neural Networks, Robust Hashing, Dataset Obfuscation, MobileNet, ResNet, DenseNet, Adversarial Attacks, Cybersecurity.
- 10 May 2025
- Published by Springer Nature
- p. 143-167
Applying Word Embeddings and Graph Neural Networks for Effective Malware Classification
- 10 May 2025
- Published by Springer Nature
- p. 143-167
Applying Word Embeddings and Graph Neural Networks for Effective Malware Classification