Oudoum Ali Houmed

Applied research on constraint-guided virtual patient generation for longitudinal clinical datasets: conditional generation pipelines with explicit filtering for statistical fidelity and temporal coherence, and evaluation of downstream model safety with a focus on calibration and generalization under distribution shift.

Fellowship project under Krystal Jackson (UC Berkeley CLTC): Can We Trust Deception Monitors for AI Agents? A Robustness-Gap Protocol and the Limits of Adaptive Evasion — on Llama-3.3-70B with the published Apollo probe, an adversary-budget ladder (b₀–b₄), and a retention gate before reporting detection drop. Findings feed technical standards, safety controls, and policy recommendations. Completed the intensive 5-week ARENA curriculum.

Investigated the vulnerability of medical diagnostic architectures to adversarial perturbations using gradient-based optimization. Implemented white-box attacks (FGSM, PGD) to quantify the robustness of tumor detection models; contributed empirical findings toward a forthcoming manuscript (Adversarial Threats to Safety-Critical Medical AI).

Achieved an internship grade of 95%. Detected and analyzed RDP brute-force attacks using Wazuh SIEM; performed deep-packet inspection with Wireshark and web-app pen-testing with Burp Suite; conducted malware scanning, static/dynamic reverse engineering, and forensic analysis with MDR tools.

Co-developed Cyber Inspector, a WAF-integrated ML system for malicious query detection. Managed the full project lifecycle: data collection, preprocessing, model training (SVM, RF), and deployment (Detection of Malicious Web Queries with Machine Learning).