Postdoc research project

1. Multimodal Machine Learning for ADHD and ASD Classification Using Computer Vision and Crowdsourced Annotations

I developed AI-powered multimodal models for diagnosing individuals with ADHD and ASD using computer vision techniques. This involves leveraging GuessWhat? videos and extracting diverse feature modalities, including facial expressions, activity recognition, speech patterns, body language, and eye gaze. By applying various AI models, it aims to evaluate confidence intervals and enhance precision in healthcare diagnostics. We integrate Human-in-the-Loop Machine Learning with crowdsourced annotations via Amazon AWS MTurk to further improve model accuracy. This approach replaces low-confidence AI predictions with high-quality human annotations, refining model performance. A comparative analysis of AI-generated and human-labeled annotations provides insights into optimizing AI-driven interventions with human expertise for precision diagnostics. 

Research projects during PhD

2. Research on Safe Access Method of Multi-party Synthesized Gene Information 

I orchestrated the development of a secure method for accessing and managing multi-party synthesized gene data. I transformed digital information into DNA sequences by innovating encoding and decoding techniques. This involved constructing a bio-coding-constraints computation model and optimizing DNA codes to enhance storage density. We successfully synthesized, sequenced, and stored data in DNA, culminating in a robust and efficient DNA-based data storage system. This project fortified my computational codec skills and problem-solving abilities, resulting in a pioneering solution with far-reaching implications for secure data storage in DNA and retrieval.

3. Method of Social Network Summarization for Information Communication Understanding

An initiative aimed to enhance communication understanding in social networks. It employed innovative techniques for detecting overlapping communities using symmetric nonnegative matrix factorization in ego-splitting networks, enriching the grasp of network dynamics. GAWA, a hybrid sentiment classification method, enabled nuanced sentiment analysis. Through Twitter data, a hybrid recommendation model for social network services was developed. Integrating a novel word-embedding approach and an LSTM-CNN model enabled real-time sentiment analysis. Pioneering a hybrid sentiment method correlating with COVID-19 tweets unveiled valuable sentiment trends. The LSTM-based model showed proficiency in finance by predicting stock market shifts during unexpected incidents, highlighting dynamic forecasting. The project demonstrated adeptness across domains, a testament to the dedication to unraveling intricate communication patterns in social networks.

4. Security control and credible enhancement management system based on large-scale dynamic infrastructure 

The project aimed to reinforce security control and credibility within large-scale dynamic infrastructures. Notable contributions include an adept detection approach for phishing websites, utilizing URL and HTML features through machine and deep learning models. A feature-based robust method was introduced for detecting abnormal contracts within the Ethereum blockchain. Additionally, efforts were dedicated to merging IoT and Blockchain to ensure secure supply chain transactions. A pioneering secure and distributed architecture was developed, specifically tailored for privacy-preserving healthcare systems. The project is a testament to our commitment to advancing security solutions within dynamic environments.