Thesis (M.Sc)-Cairo University, 2025.

Bibliography: pages 81-88.

Detecting artificially inflated SMS traffic is crucial in telecommunications to
maintain network integrity amid rising fraudulent activities like SMS spamming.
This thesis explores machine learning techniques for identifying such anomalies,
comparing four models: RandomForestClassifier, GradientBoosting, Support
Vector Machine (SVM), and Naive Bayes (GaussianNB). Utilizing a dataset of
12,500 SMS instances 10,000 labeled as artificially inflated and 2,500 as normal.
The study examines features including sender and recipient numbers, mobile country
code (MCC), mobile network code (MNC), country, and network, all preprocessed
for modeling.
The primary goal is to evaluate each model's effectiveness in distinguishing
artificially inflated SMS traffic from normal messages, focusing on metrics like
precision, recall, and F1-scores. RandomForestClassifier achieved high accuracy
with a precision of 0.91 for AIT and 1.00 for normal traffic; however, it had a lower
recall for normal traffic (0.59), indicating challenges in identifying certain normal
instances. The SVM model, despite its proficiency in high-dimensional spaces,
showed lower precision and recall for normal traffic, suggesting difficulties in
classifying non-inflated messages. GaussianNB performed poorly for normal traffic,
with precision, recall, and F1-scores near zero, indicating its limitations with
complex or imbalanced datasets. In contrast, GradientBoostingClassifier delivered
promising results, with performance metrics comparable to RandomForestClassifier,
demonstrating its effectiveness in detecting artificially inflated SMS traffic.
The thesis analyzes each model's strengths and weaknesses, identifying contexts
where one may be more suitable than others. RandomForestClassifier and
GradientBoostingClassifier emerged as the most reliable for detecting artificial
inflation, with GradientBoosting showing greater resilience with imbalanced
datasets. SVM's performance was hindered by its inability to effectively identify


v

normal traffic in this scenario, while GaussianNB's simplifying assumptions limited
its utility.
This comprehensive evaluation underscores the importance of selecting appropriate
machine learning models based on data characteristics and task requirements. While
advanced models like RandomForest and GradientBoosting offer superior
performance in this domain, simpler models such as Naive Bayes or SVM may still
be applicable in specific scenarios with proper data preparation. Future research
could explore other algorithms, hyperparameter optimization, and incorporation of
domain-specific features to enhance detection accuracy.
In conclusion, the study confirms that machine learning effectively identifies
artificially inflated SMS traffic, providing telecom operators and fraud detection
systems with viable solutions to mitigate SMS-based fraud risks. Comparative
analysis offers valuable insights for selecting suitable techniques based on data
availability, computational resources, and specific detection system objectives. .

يُعد تضخم حركة الرسائل النصية القصيرة (SMS) بشكلٍ اصطناعي من أخطر التحديات التي تواجه شركات الاتصالات، لما له من آثار مالية وتشغيلية سلبية. تهدف هذه الدراسة إلى تطوير نموذج تنبؤي يعتمد على تقنيات تعلم الآلة لاكتشاف الرسائل النصية المصطنعة التي تُستخدم غالبًا في هجمات احتيالية مثل إساءة استخدام رموز التحقق (OTP).
تركز الدراسة على تحليل بيانات حقيقية مكونة من 12,500 رسالة قصيرة، منها 10,000 تم تصنيفها كحركة مضخّمة اصطناعيًا و2,500 رسالة طبيعية. شملت الدراسة استخراج وتحليل عدد من الخصائص (features) المرتبطة بالمرسل، المستلم، ومعلومات الشبكة مثل رمز الدولة (MCC) ومُشغّل الشبكة (MNC)، وتمت معالجة البيانات باستخدام تقنيات مناسبة للتشفير والتنظيف.
تم بناء وتقييم أربعة نماذج تصنيف باستخدام خوارزميات تعلم الآلة: Random Forest، Gradient Boosting، Support Vector Machine (SVM)، وGaussian Naive Bayes. وأظهرت النتائج أن نماذج Random Forest وGradient Boosting حققت أعلى أداء، حيث تفوقت في الكشف عن الرسائل الاصطناعية بدقة عالية، بينما أظهرت خوارزميات SVM وNaive Bayes أداءً أقل خاصة في التعامل مع البيانات غير المتوازنة.
تقدم هذه الدراسة مساهمة عملية في تعزيز أمان شبكات الاتصالات، من خلال تقديم إطار عمل قابل للتطبيق يُمكّن مزودي الخدمة من التعرف المبكر على التضخم الاصطناعي للرسائل، مما يدعم تحسين الكفاءة التشغيلية وتقليل الفاقد المالي.

Issues also as CD.

Text in English and abstract in Arabic & English.