Projects

This project creates high-density LiDAR areas by leveraging high-reflectivity surfaces, ensuring robust and effective trigger placement. 

For experiments, configured and deployed camera-LiDAR sensors on autonomous vehicles to collect data for security and robustness analysis.

In this project, we developed a genetic optimization framework to refine malicious intent prompts and passages for stealth and effectiveness.

Achieved a 45% improvement in attack accuracy over existing methods.

With the increasing adoption of camera-LiDAR fusion for high-fidelity 3D object detection, the robustness of these systems against adversarial threats is a growing concern.

We introduce BadFusion, an innovative 2D-oriented backdoor attack that preserves trigger effectiveness across the entire fusion process, achieving significantly higher attack success rates compared to existing 2D-oriented attacks on 3D object detection models.

Despite advancements in deep neural networks, object detection in adverse weather remains a significant challenge due to sensor degradation under extreme conditions. 

To address adverse weather detection challenges, our paper introduces a Global-Local Attention (GLA) framework that adaptively fuses multiple sensor streams—camera, gated, and LiDAR data—at two fusion stages:

• Early-stage fusion via a local attention network, capturing localized feature variations.

• Late-stage fusion via a global attention network, assigning higher importance to the most reliable modality under the current weather condition.

Skin cancer is a growing global health concern, with 123,000 melanoma and 3,000,000 non-melanoma cases reported annually worldwide. Excessive exposure to ultraviolet rays is a major risk factor.

This study presents an efficient deep learning-based skin cancer classification model to assist dermatologists in critical decision-making for early-stage skin cancer detection, achieving an accuracy of 83.1%.

Diabetic Retinopathy (DR) is a leading cause of vision loss, affecting over 100 million individuals worldwide, with cases expected to rise in the coming decades.

To address this, we developed a deep learning-based diagnostic tool for early detection of Diabetic Retinopathy, automating disease classification and achieving 96% accuracy. Improved detection accuracy by 35%, surpassing previous studies and outperforming traditional screening methods.