● Research Showcase

Hybrid Vulnerability Detection:
ML + Policy-as-Code

CodeBERT (ML) and Semgrep (PaC) combined with score normalization and a minimum policy weight — evaluated on the DiverseVul C/C++ dataset.

ML-Only (CodeBERT)
01
Probabilistic detection
Outputs a continuous confidence score learned from 260k+ labeled functions
02
Catches semantic patterns
Detects subtle issues like timing side-channels or logic errors — no explicit rules needed
03
Low precision on its own
P=0.127, F1=0.186 on test set; needs policy to improve signal quality
PaC-Only (Semgrep)
01
Deterministic rules
Every match is explicit and auditable; rules target top CWEs (CWE-119, 416, 476, 190…)
02
High precision, low recall
P=0.206 but Recall=0.033; fires on a small fraction of the test set
03
Misses semantic bugs
Timing attacks, logic errors, and non-trivial UAF patterns are invisible to rules
Hybrid Pipeline
📄
Input Code
C/C++ function (DiverseVul)
→
🧠
ML (CodeBERT)
P(vulnerable) → [0,1]
→
🔍
PaC (Semgrep)
Findings → score [0,1]
→
⊕
Hybrid Fusion
α·norm(ML) + β·norm(PaC)
→
🚦
Decision
Approve / Review / Block
Tuned Config
α (ML weight) 0.75
β (PaC weight) 0.25
t_block 0.40
t_review 0.10
min_pac_weight 0.20
Scores normalized to [0,1] using validation set min/max before fusion. Ensures both channels are on a comparable scale.
Code Scenarios — Where Hybrid Wins
All Scenarios at a Glance
Illustrative scores derived from real Phase 3 config (α=0.75, β=0.25, t_block=0.4, t_review=0.1, normalization applied).
Scenario CWE ML raw ML decision PaC score PaC decision Hybrid risk Hybrid decision Story
Live Analysis — Try It
API URL Local API URL to enable live analysis (run the API from the repo with USE_REAL_ML=1 uvicorn api.main:app --reload)
your_code.c