Final Year Project
Final Year Project
Presented By
Aditya Chug 1NT21EC004
Ruchi Yadav 1NT21EC117
Shreyas Somanache 1NT21EC148
Project Supervisor
Prof. Prajna.K.B
Associate Professor
Dept. of Electronics and Communication Engineering
Nitte Meenakshi Institute of Technology
Yelahanka, Bangalore-560064
Dept. of ECE, NMIT, Bengaluru
Presentation Outline
• Introduction
• Literature Survey
• Project Objectives
• Tools Required
• Architecture
• Mathematical modelling design approach
• Proposed System Methodology
• Expected Outcomes
• References
• Provide a responsive and user-friendly interface with form validation and instant fraud detection feedback.
• To address the challenge of imbalanced datasets by implementing techniques such as oversampling, under-
sampling, and anomaly detection algorithms suited for rare-event detection.
• To evaluate the system’s robustness against adversarial attacks and ensure it can withstand attempts by
fraudsters to evade detection.
• Provide real-time alerts and detailed reports on flagged transactions via the interface.
2. Preprocessing Layer
• Purpose: Clean and prepare data for analysis.
• Technology: Python (Pandas, NumPy, Scikit-learn).
• Details:
• Handle missing values using imputation techniques.
• Scale numeric data for SVM compatibility (e.g., Min-Max Scaling).
• Perform one-hot encoding for categorical variables.
• Balance the dataset to address class imbalance (e.g., SMOTE).
Dept. of ECE, NMIT, Bengaluru
Basic Architecture
3. Modeling Layer
• Purpose: Apply machine learning models for fraud detection.
• Algorithms Used:
• Extra Tree Classifier: Randomized tree splits for fast computation and reduced overfitting.
• Support Vector Machine (SVM): Separates classes using hyperplanes, effective for small but complex
datasets.
• XGBoost: Gradient boosting framework with regularization for high accuracy.
• Technology: Scikit-learn, XG Boost library
4. Decision Layer
• Purpose: Output fraud probabilities for transactions.
• Details:
• Ensemble Approach: Combines predictions from Extra Tree, SVM, and XGBoost.
• Decision Threshold: Assigns "fraudulent" or "legitimate" labels based on a set confidence level.
• Technology: Python Flask/Django for REST API, model serialization with joblib or pickle.
• Evaluation Metrics:
• Accuracy
• Precision
• Recall (Sensitivity)
2. Speeds Up Convergence:
• Helps optimization algorithms (like gradient descent) converge faster by eliminating bias
introduced by differing feature scales.
• Goal: Build multiple decision trees using random splits and aggregate results.
• Mathematical Representation:
f(x)=
where:
• T: Number of decision trees
• ft(x): Prediction from tree t
• Goal: Find the optimal hyperplane that separates fraudulent and non-fraudulent
transactions.
• Hyperplane Equation:
w⋅x + b = 0
where:
• w: Weight vector
• x: Feature vector
• b: Bias term
where:
• TP: True Positive
• FP: False Positive
• FN: False Negative
• TN: True Negative
Pre-processed
Initiate Dataset
Transaction
Classification
Techniques