Research
Security testing at Georgia Tech · a NeurIPS 2023 workshop paper
Current
In progress
LLM-driven security testing
Systems Software & Security Lab, Georgia Tech. Advisor: Prof. Taesoo Kim.
The framework uses shallow bugs as coverage signals to explore deeper execution paths, then patches what it finds automatically and evaluates the patches. It runs across databases, networking tools and bootloaders.
The framework builds on ATLANTIS, an AI-driven vulnerability discovery system and winner of DARPA’s AI Cyber Challenge (AIxCC); I wired it into Docker-based CI/CD, using call graphs to trace callers and dependencies of changed functions, generate targeted fuzzing harnesses and compare coverage of new code against existing harnesses.
- 11
- 5
Publication
arXiv:2309.11611- North Africans in Machine Learning (NAML) workshop, NeurIPS 2023. Accepted poster.
- Co-author and model evaluation task lead, through the Omdena Algiers Local Chapter.
- The benchmark favored DzaraShield at 0.87 F1, against multilingual E5 at 0.80 and SBERT at 0.77, on 13.5K Algerian-dialect texts.
- Cited in six publications.
All 10 models from Table 1. Model F1 · bar from 0 to 1 DzaraShield 0.87 DZiriBERT FT PEFT+LoRA 0.84 AraT5v2-HateDetect 0.83 Dziribert-FT-HEAD 0.81 Multilingual-E5-base FT 0.80 sbert-distill-multilingual FT 0.77 Bidirect LSTM 0.68 LSTM 0.67 gzip + KNN 0.60 LinearSVC 0.9 (Class 0) · 0.48 (Class 1) † Table 1 reports LinearSVC’s F1 per class only, so it has no bar. The paper labels class 0 non-hate and class 1 hate.