Benhur Tekeste
AI Security • Adversarial Machine Learning • Network and
Systems Security
Research Interests
AI security and adversarial machine learning; realistic robustness
evaluation of ML-based security systems; security of multi-agent
(agentic) AI systems; AI red teaming; ML for network and systems
security.
Education
Khalifa UniversityJan 2019 – Dec 2023
B.Sc. in Computer Engineering • GPA: 3.86/4.00Abu Dhabi, UAE
Publications
Towards Realistic Adversarial Robustness Evaluation for
Intrusion Detection in Industrial Control Systems
Benhur Tekeste, et al. IEEE International
Symposium on Network Computing and Applications (NCA), 2026.
Accepted. (First author)
Shows that constraint-blind evaluation overestimates the
adversarial vulnerability of ML-based ICS intrusion detectors by
up to 80–99 percentage points, by isolating in-loop vs.
post-hoc constraint enforcement (JSMA, EAD, C&W, PGD; SWaT
and WADI testbeds).
Adversarial Machine Learning: A 20-Year Survey of Attacks,
Defenses, and Standards
Benhur Tekeste, et al. IEEE Access, 2025.
(First author)
Systematizes two decades of adversarial ML research, covering
attack methodologies, defense mechanisms, and the evolving
standards landscape for ML security.
Research Experience
Research Assistant – Cybersecurity and AIJul 2025 – Present
Rochester Institute of Technology DubaiDubai, UAE
- Conduct research on securing AI systems for smart-city applications
- Research outputs include first-author publications on adversarial ML [J1] and realistic adversarial robustness evaluation for ML-based intrusion detection [C1]
Research Assistant – ML-Based Routing for Wireless Ad Hoc
NetworksFeb 2024 – Jan 2025
Khalifa UniversityAbu Dhabi, UAE
- Developed QoS-OLSR, an ML-based routing protocol building on OLSR, to optimize network performance in wireless ad hoc networks
- Built a high-fidelity Mininet-WiFi emulation testbed to validate protocol behavior under realistic wireless constraints
- Integrated QoS-OLSR with Technology Innovation Institute (TII) Comm Sleeve devices running Hardened OS to strengthen the security of the deployed routing stack
- Extended the protocol with IPv6 support, achieving compliance with IPv6 standards
- Mentored undergraduate capstone students implementing a variant of the ML-based QoS-OLSR protocol
Academic Service
Artifact Evaluation Committee Member, Network and Distributed
System Security (NDSS) Symposium2027
Selected Projects
Membership Privacy of LoRA Fine-Tuning and DP-LoRA in Open-Weight
LLMs
| PyTorch, PEFT/LoRA, Opacus
- Designed a randomised, multi-seed study of training-data leakage in Qwen2.5-1.5B/3B models fine-tuned with LoRA (130 models). Members were drawn at random from one deduplicated pool, and a text-only check confirmed no member/non-member shift (AUC 0.50)
- Implemented 9 membership-inference attacks, from loss thresholds to offline LiRA with recipe-matched reference models, plus DP-LoRA from scratch (Poisson sampling, per-record clipping, RDP accounting cross-checked against two independent accountants)
- Found that the model with the best held-out loss already leaks: LiRA identifies about 60% of training records at 1% FPR, while loss attacks and verbatim extraction look harmless. Calibrating with the base model can understate this risk sharply (0.18 vs 0.99 TPR at 3B)
- Showed that DP-LoRA at ε ≤ 8 cuts leakage to near chance while keeping 51% (1.5B) and 86% (3B) of the utility gain. A noise-free control showed that the noise, not the clipping, provides the protection
Multi-Agentic AI System for Financial Analysis
| CrewAI, LangGraph, Multi-Agent Building
- Built a CrewAI-based collaboration of four specialized agents, namely a data analyst, strategy developer, trade advisor, and risk advisor
- The agents work through a hierarchical process to analyze a stock and produce a trading recommendation
- Agents use live web search and site-scraping tools to gather market data, and take a stock ticker, capital, risk tolerance, and trading preferences as input
Awards and Recognition
2nd Place, Cyber Energy Drill, ADIPEC2023
3rd Place, Khalifa University Programming Contest2023
Sustainability Expert Award, GITEX AI InnovateFest2023
Distinguished Peer Tutor, Khalifa University2023
Inducted to Golden Key International Honour Society2022
Certifications
Certified Red Team Operator (CRTO), Zero-Point SecurityApr 2025
eLearnSecurity Junior Penetration Tester (eJPT), INE
SecurityJun 2024
Technical Skills
- Security Domains:
- Adversarial ML, network security, penetration testing and red teaming, malware and firmware analysis, log forensics
- Programming:
- Python, C, Java, Bash, PowerShell, SQL
- ML / AI:
- TensorFlow, Hugging Face Transformers, LangChain, Pandas, Matplotlib, Gradio
- Networking:
- Mininet-WiFi, OLSR, IPv6, Wireshark
- Security Tools:
- Ghidra, Burp Suite, Metasploit, Cobalt Strike, PowerShell Empire, Nessus, Splunk
- Systems:
- Linux, Windows, Docker, AWS, Git, Flask, LaTeX