Peer-Reviewed Research & Active Manuscripts

Publications in Trustworthy Machine Learning and Cybersecurity

My research examines how probabilistic models, generative systems, uncertainty, reproducible experimentation, and downstream evaluation can support dependable cybersecurity and scientific-computing applications.

Synthetic Cybersecurity Data Malware Classification Generative Models Bayesian Deep Learning Uncertainty Estimation Scientific Data Recovery Reproducible Research

View Google Scholar Explore Research Research Contact

Published Peer-Reviewed Work

Research that has completed peer review and is available through established scholarly publishers.

2026 Published Springer Conference Paper

Towards Reliable and Interference-Aware CSI Feedback with Bayesian Neural Network

Ifiok Udoidiok, Bruno Fonkeng, Jielun Zhang, and Fuhao Li

Proceedings of the International Symposium on Intelligent Computing and Networking 2025 · Lecture Notes in Networks and Systems, Volume 1698 · Springer

Presents a Bayesian neural-network framework for reliable channel-state-information feedback. The framework combines CSI compression and reconstruction with uncertainty estimation to detect interference and improve feedback reliability under non-ideal transmission conditions.

Research contribution spanning Bayesian deep learning, uncertainty-aware reconstruction, interference detection, experimental evaluation, and technical manuscript development.

Bayesian Neural Networks Uncertainty Estimation Interference Detection Wireless Systems

Springer Paper ↗ DOI ↗

2024 Published IEEE Conference Paper

Masked Autoencoder for Data Recovery in Polymer Research: Mitigating Data Integrity Threats

Bruno Saidou Fonkeng, Fuhao Li, Binglin Sui, and Jielun Zhang

2024 Cyber Awareness and Research Symposium · IEEE

Introduces an autoencoder-based recovery method that uses random masks to reconstruct missing or corrupted polymer-research data. The approach was evaluated on cloud-point-temperature datasets and demonstrated stronger recovery performance than conventional statistical-imputation methods.

First-author work connecting machine learning, scientific data recovery, polymer research, and protection against data-integrity loss caused by failure, deletion, or cyberattack.

First Author Autoencoders Data Recovery Data Integrity Scientific Computing

IEEE Xplore ↗ DOI ↗

Submitted Research

Completed scholarly work formally submitted for peer review. Submission does not imply acceptance or publication.

2026 Submitted IEEE Conference Submission

GenCyberSynth: A Unified Comparative Benchmark for Synthetic Malware Images

Bruno Fonkeng, Regine Khun, Ankit Tiwari, Jielun Zhang, and Fuhao Li

Submitted to the 2026 IEEE Global Communications Conference, Communication and Information Systems Security Symposium

Presents GenCyberSynth as a unified and reproducible benchmark for comparing synthetic malware-image generators under controlled conditions. The framework standardizes data preparation, train–synth–evaluate workflows, random seeds, synthetic budgets, leakage-safe testing, real-only baselines, downstream utility metrics, and structured result aggregation.

First-author benchmark work integrating seven generator families, multiple malware-image datasets, reproducible HPC workflows, proxy generative diagnostics, and downstream classification evaluation.

First Author GenCyberSynth Benchmarking Synthetic Malware Images Reproducibility Downstream Utility

Completed Manuscripts Ready for Submission

Completed research manuscripts in the GenCyberSynth publication series. These works are ready for submission but are not presented as submitted, accepted, or published.

Paper 2 Submission Ready Conference Manuscript

Conditional Generation Done Right: External Conditioning Audits for Synthetic Malware Images

Bruno Fonkeng, Regine Khun, Jielun Zhang, and Fuhao Li

Completed manuscript · Preparing for submission

Introduces an external conditioning audit for class-conditional malware-image generators. Requested generation labels are compared against predictions from independent real-only classifiers to measure conditioning accuracy, label leakage, confusion patterns, and predicted-class collapse.

Shows that internal conditioning mechanisms and downstream augmentation gains do not necessarily prove requested-label fidelity, motivating independent conditioning audits for class-specific synthetic cybersecurity data.

External Auditing Conditional GANs Class Fidelity Label Leakage Failure Analysis

Paper 3 Submission Ready Conference Manuscript

When Does Synthetic Data Help Malware Classification? Utility, Class Imbalance, and Failure Modes Across Augmentation Regimes

Bruno Fonkeng, Regine Khun, Jielun Zhang, and Fuhao Li

Completed manuscript · Preparing for submission

Studies when GAN-, VAE-, and diffusion-generated malware images improve downstream classification and when they harm it. The paper compares balanced all-class augmentation with a controlled minority-heavy regime across multiple budgets and random seeds.

Demonstrates that synthetic augmentation is regime-dependent: value depends on generator family, budget, class structure, augmentation policy, and downstream metric rather than synthetic sample count alone.

Downstream Utility Class Imbalance GAN VAE Diffusion Failure Modes

Paper 4 Submission Ready Conference Manuscript

Selective Synthetic Data Policies for Malware Classification: Allocation, Filtering, and Acceptance Rules Under Utility Constraints

Bruno Fonkeng, Regine Khun, Jielun Zhang, and Fuhao Li

Completed manuscript · Preparing for submission

Investigates which generated malware samples should be retained, filtered, ranked, or allocated before downstream classifier training. The work compares fixed budgets, keep-all augmentation, strict-confidence filtering, Top-k selection, and adaptive class-repair allocation.

Treats synthetic augmentation as a policy-controlled intervention and evaluates whether reduced or adaptively allocated synthetic sets can preserve utility under storage, training-cost, and acceptance constraints.

Selective Augmentation Sample Filtering Top-k Selection Adaptive Allocation Utility Constraints

GenCyberSynth Research Progression

The four-paper sequence moves from benchmark construction to conditional reliability, downstream utility, and policy-controlled synthetic-data selection.

Paper 1

Unified Benchmark

Establishes the reproducible evaluation framework, controlled baselines, generator comparison, and utility-first protocol.

Paper 2

Conditioning Audits

Tests whether class-conditioned samples are externally recognized as the malware classes they were requested to represent.

Paper 3

Utility and Failure Modes

Determines when synthetic augmentation helps, when it hurts, and how results change across budgets and class-imbalance regimes.

Paper 4

Selective Policies

Evaluates filtering, ranking, acceptance, and allocation policies for using synthetic samples under practical utility constraints.

Publication statuses are intentionally precise. “Submitted” indicates formal submission for peer review. “Submission ready” indicates a completed manuscript that has not yet been formally submitted. Neither status implies acceptance or publication.

Research Built Around Defensible Evidence

My publication work connects probabilistic modeling, synthetic cybersecurity data, uncertainty, reproducibility, downstream utility, failure analysis, and the engineering infrastructure required to produce trustworthy scientific conclusions.

Explore Research View Projects Discuss the Work