Karim Razgallah — PhD applicant in systems security and program analysisKarimRazgallah

Showing all 3 projects.

1. [01]

   MAM: Multimodal Agentic Memory

   Karim Razgallah, Vijay K. Madisetti

   venueyearkind

• 01

  Generalization in binary program analysis

  A model that identifies a patched function in GCC -O0 x86 binaries often fails on the same function built with Clang -O2 for ARM. The failure is not subtle: inlining, register allocation, and instruction selection rewrite the very code the model learned to read. I want to work out which representations survive that rewriting, and which only appeared to work because training and evaluation drew from the same build environment. The concrete target is an evaluation protocol where held-out means a held-out compiler, not a held-out sample.

• 02

  Calibrated abstention and trustworthy security tooling

  A tool that is right ninety percent of the time is not usable if nobody can tell which ninety percent. An analyst triaging a vulnerability needs the tool to separate cases it has real evidence for from cases where it is extrapolating. I want to study abstention that is calibrated rather than heuristic: a confidence score that means the same thing across compilers and architectures, and a rejection threshold an operator can set against a stated error budget.

• 03

  TODO: third research direction

  one short paragraph naming a specific problem. Candidates that follow from the work above: patch propagation across forked and vendored codebases, provenance recovery from stripped binaries, or evaluating whether a shipped dependency matches the source it claims to be built from.