P02Telecom · Cybersecurity2024UPEC
P02

SimBox Fraud Detection

Imbalanced classification + metaheuristic search

CASE STUDY · FINAL-YEAR PROJECT (PFE)

CLASS RATIO

≈ 1 : 200

OPTIMISATION

Metaheuristic

TARGET

False-positive rate

DOMAIN

International voice

STACK //scikit-learnMetaheuristicsEDAImbalanced learning

STEP / 01

Brief

SimBox fraud reroutes international calls through grey-market hardware, costing operators serious revenue. The detector has to find a moving target in a wildly imbalanced stream.

STEP / 02

Approach

Build a feature space the model can actually learn from.

  • 01Exploratory analysis to surface call-pattern fingerprints
  • 02Engineered features around call duration, IMEI churn, geo-rotation
  • 03Imbalanced-class strategies (resampling, class-weighting)
  • 04Metaheuristic search over hyperparameters and feature subsets

STEP / 03

Model

Classic supervised classifiers tuned by metaheuristic optimisation. The win was a tighter false-positive rate without giving up recall on a constantly evolving fraud pattern.

STEP / 04

Friction points

Why this was not a textbook problem.

  • 01Severe class imbalance with drifting prevalence
  • 02Adversarial fraud patterns that mutate week-over-week
  • 03Operator-side cost asymmetry between FP and FN

STEP / 05

Outcome

Improved detection rate with materially fewer false alarms — a model robust enough to ride drift without retraining every week.

Strategic detection work for a domain where every false positive costs a real customer.

NEXT CASE / P03

Geo Customer Dashboard