Detailed Projects
In-depth look at my key projects and research interests
Multi-Modal Facial Aesthetics Assessment using Deep Learning and Geometric Modeling
This project investigates the problem of objective facial aesthetics assessment by formulating it as a multi-modal learning task that jointly models facial texture, symmetry, and geometric structure. Unlike purely data-driven cosmetic scoring systems, the proposed framework integrates deep convolutional representations with explicit geometric constraints derived from facial landmark topology, enabling interpretable and reproducible aesthetic evaluation.
Facial Landmark Geometry
Deep Feature Fusion
Methodology & Technical Contributions:
- Extraction of anatomical facial landmarks and geometric ratios using morphometric analysis from SCUT-FBP5500 Dataset
- Deep CNN-based feature learning for texture, symmetry, and proportional consistency
- Multi-modal fusion architecture (EfficientNet_V2, MLP & CNN) combining geometric descriptors with learned embeddings
- Regression-based aesthetic scoring with correlation-based validation metrics
Research Significance:
The framework bridges biomedical image analysis and human-centered AI by introducing explainable structure-aware modeling, with applications in orthodontics, cosmetic planning, medical imaging, and human–computer interaction.
Results & Interpretation
The proposed facial aesthetics assessment model achieves a final MAE of 0.214 and RMSE of 0.282 on the validation set, indicating accurate regression performance on subjective aesthetic scores. Performance improves steadily across training epochs, converging to low error values without overfitting. These results demonstrate effective feature learning and reliable generalization on the SCUT-FBP5500 dataset.
Prompt Injection Detection using Classical NLP and Transformer Models
This project investigates the vulnerability of large language models to prompt injection attacks by designing a structured detection pipeline that combines classical NLP feature engineering with modern transformer-based representations. The work frames prompt injection as a semantic and syntactic anomaly detection problem.
Classical features such as n-gram frequency, POS distributions, and semantic similarity scores are integrated with contextual embeddings from transformer encoders. This hybrid design improves detection robustness across diverse attack patterns while maintaining interpretability.
Linguistic Feature Engineering
Transformer-Based Detection
Core Methodology:
- Benign and malicious prompts are merged, cleaned, and labeled
- Prompts are represented using TF-IDF and transformer tokenization
- Logistic Regression and DistilBERT models are trained
- Evaluation across adversarial and benign prompt distributions
Results & Interpretation
The model demonstrates strong and reliable performance, achieving 99.5% test accuracy with an F1-score of 0.992 and an AUC of 0.997, indicating highly effective prompt-injection detection. Results remain consistent across training, validation, and testing splits, showing robust generalization to unseen and adversarial prompts. This stability and low error rate highlight the approach as a practical, deployment-ready safeguard for real-world LLM systems.
Numerical Solution of Diode Circuits using Hybrid Root-Finding Methods
This project addresses the numerical instability and convergence limitations of classical methods used to solve nonlinear diode circuit equations. A hybrid root-finding framework is developed by combining the robustness of the Newton Raphson method with the rapid convergence properties of Newton–Raphson and Secant Method, enabling stable and efficient computation of diode operating points.
Nonlinear I–V Modeling
Convergence Analysis
Methodology & Analysis:
- Formulation of diode circuits using exponential Shockley equations
- Hybrid numerical solver design for guaranteed convergence
- Error tolerance and iteration complexity analysis
- Comparative validation against standalone numerical techniques
Results & Interpretation
The hybrid approach converged in 5 iterations, reducing computational error as compared to standalone methods such as Newton Raphson, Secand Method or Bisection Method which converge to the solution after 49, 62, 31 iterations respectively. The solution remained stable across all tested operating points.
Inter-Microcontroller Communication System using AVR Architecture
This project implements a reliable communication protocol between two AVR microcontrollers, focusing on low-level data transfer, synchronization, and timing accuracy. The system is designed at the register and instruction level, emphasizing deterministic behavior and hardware-aware software design.
Data Transmission
Timing Synchronization
System Design Highlights:
- Assembly-level implementation on ATmega architecture
- Register-level I/O configuration and timing control
- Error handling through controlled handshaking
- Validation using logic-level signal observation
Results & Interpretation
Reliable bidirectional communication was achieved between two ATMEGA 32 microcontrollers. The transmitter sent digits 1–9 and 0, and the receiver correctly stored and displayed them in reverse order.
Adversarial Attacks on Image Classification using FGSM
This project experimentally evaluates the vulnerability of lightweight CNN architectures to adversarial perturbations using the Fast Gradient Sign Method (FGSM). MobileNetV2 is selected due to its relevance in embedded and real-time systems.
The study quantifies accuracy degradation under varying perturbation magnitudes and analyzes feature-space sensitivity to adversarial noise, highlighting risks in safety-critical vision applications.
Analysis Focus:
Procedure
Gradient-based adversarial perturbation generation
Validate
Model robustness evaluation across epsilon ranges
Results and Interpretations
The baseline model initially predicted Labrador Retriever with 41.82% confidence. As the perturbation magnitude (ε) increased, confidence degraded and classification labels shifted (e.g., Saluki at ε = 0.01 with 13%, Weimaraner with ~15% at ε = 0.1), demonstrating both confidence collapse and label poisoning under adversarial noise.
Breast Cancer Classification using Support Vector Machines
This project applies support vector machines to classify malignant and benign breast tumors using structured biomedical feature sets. Emphasis is placed on margin optimization and generalization performance under limited-sample conditions.
Kernel selection, feature scaling, and hyperparameter tuning are systematically evaluated to maximize diagnostic sensitivity while minimizing false negatives.
Biomedical Feature Space
Classification Performance
Methodology & Technical Contributions:
- Preprocessed diagnostic features using normalization to ensure scale-invariant margin optimization
- Trained a Support Vector Machine with an optimal kernel to maximize class separability between malignant and benign samples
- Evaluated model performance using confusion matrix analysis and accuracy metrics
- Validated generalization through controlled train–test partitioning
Results & Interpretation:
The SVM classifier demonstrated strong discriminative capability, achieving an overall accuracy of 95.12%. The confusion matrix indicates high true-positive and true-negative rates with minimal misclassification, reflecting effective margin optimization and robust separation between benign and malignant classes. These results validate SVM suitability for reliable, low-variance clinical decision support in breast cancer diagnosis.
Audio Signal Processing for Noise Removal and Feature Extraction
This project focuses on classical and modern signal processing techniques for audio denoising and feature extraction. The system processes raw audio signals to enhance signal-to-noise ratio while preserving perceptually relevant information.
Time–frequency analysis, digital filtering, and spectral feature extraction are combined to support downstream machine learning and speech-processing applications.
Time–Frequency Analysis
Noise Filtering Pipeline
DSP Techniques:
- Audio signals were recorded at a fixed sampling rate using MATLAB’s audiorecorder
- Synthetic noise was generated and analyzed alongside the clean signal in the frequency domain
- The noisy signal was constructed by superimposing random noise onto the original audio and re-examined via spectral analysis
- A Butterworth digital filter was designed and applied to suppress noise components while preserving speech-relevant frequencies