Publications & Research Output.

AbstractThis thesis examines deepfake susceptibility and synthetic media perception among older adults (55+) and adolescents (11-18). Using a participatory design approach that combines interviews and surveys, it captures how these groups perceive AI-generated content. Insights from participants informed a trained vision model and a mobile application that acts as an interactive AI-awareness tool for marginal communities, analysed through Computer Vision, Human-Computer Interaction, and Social Computing perspectives.

AbstractNutriSight combines YOLO26n-Seg detection and segmentation, a ViT-LSTM model for confidence-aware recognition refinement, and a ViT-Small model for meal-level nutrition estimation. It achieves an F1-score of 0.9007 on multi-food recognition refinement and a test MAE of 18.05 for meal-level calorie and macronutrient regression, and is deployed in a mobile application with customizable dietary profiles for allergies and health conditions.

AbstractThis study re-implements four MCP security frameworks (SAMOS, AgentBound, MCPShield, and Breaking the Protocol) and evaluates them on 299 MCP servers and 27,861 metadata entries. Experiments examine least-privilege boundaries, gateway information-flow control, metadata misalignment detection (78.38% detection rate), and protocol-level vulnerabilities, and compare reproduced results with the originally reported values.

AbstractThis paper evaluates FP32, INT8 post-training quantization (PTQ), and INT8 quantization-aware training (QAT) with TinyCNN and MobileNetV2 on CIFAR-10 and CIFAR-10-C. It introduces INT8-RPTQ, a degradation-aware PTQ calibration strategy that improves corruption robustness without full QAT retraining, and analyses latency, peak memory, throughput, power, and energy per inference on edge devices.