Workflowmalware-detection
Malware Detection in Mobile Apps
A custom 2-layer neural network built from scratch (no ML frameworks) achieving 97.16% accuracy for malware detection via Android app permission analysis.
stack
- Python
- Scikit-learn
- Pandas
- NumPy
- Matplotlib
Architecture
- Feature engineering — Android permission requests encoded as binary feature vectors
- From-scratch 2-layer network — forward propagation, backpropagation, gradient descent, and activations in NumPy, no ML frameworks
- Manual training loop — weight initialization, loss computation, and gradient updates from first principles
- He initialization — hidden-layer weights scaled by sqrt(2/n) to avoid vanishing gradients
Results
- 97.16% accuracy, 98.20% F1 on malware classification
- Fully manual forward and backward propagation
Action · Send message
Hiring for a remote AI engineering role?
Contact me