Xiao D
- Research Program Mentor
PhD at Indiana University - Bloomington
Expertise
Cryptography, Privacy-preserving AI/ML, GenAI–social media psychology research, Homomorphic Encryption
Bio
I am a researcher and educator at the intersection of cryptography, AI, and algorithm design. Since 2021, I have mentored talented high-school scholars here, guiding them from initial ideas to outcomes such as peer-reviewed publications. My mentorship covers framing impactful questions, teaching technical foundations, supervising experiments, and coaching students through writing and revision. Mentees have published in the STEM Fellowship Journal, Journal of Student Research, and International Journal of High School Research. My focus areas include Cryptography, Privacy-Preserving Machine Learning, Deep Learning, and Computer Vision. I emphasize innovative thinking, rigorous implementation, and clear communication, helping students gain both publications and confidence for future projects. Outside of research, I enjoy history, oil painting, swimming, and running.Project ideas
Ask Anything, Stay Private: An Encrypted Homework Helper
This project tackles a common school problem: routing homework questions to the right subject area without exposing what students actually wrote. Many students hesitate to post in group chats because their struggles feel private. To protect them, the system encrypts each question before processing. A lightweight multi-class classifier (Math vs Biology vs Chemistry) runs on the encrypted data. The server only sees encrypted numbers, not text. The key idea is that cryptography and machine learning can work together, enabling useful predictions while keeping students’ messages private end-to-end.
Homomorphic Encryption for Privacy-Preserving Cyberbullying Detection
This project adapts an encrypted-AI pipeline to a student-relevant challenge: detecting cyberbullying in short social-media posts without exposing private messages. Cyberbullying often hides in everyday chats—harassment, threats, or subtle insults—but many students hesitate to share messages if it means losing privacy. Here, each message is transformed into an embedding, encrypted on the client side, and sent to the server. An ML model makes its prediction entirely on ciphertext, so the server never sees raw text—only encrypted inputs and outputs. The encrypted result is returned, and only the sender holding the private key can decrypt it—showing how embeddings, encryption, FHE, and machine learning combine to provide private, real-time support.
When Pictures Speak: Linking Art and Poetry with AI
Art and writing are two different forms of expression—visual and verbal—but they often inspire each other. This project uses a multimodal AI model that can understand both images and text by placing them in the same “semantic space.” Students can explore which paintings and poems naturally align: Do impressionist artworks connect more with nature poems? Do abstract pieces link with modern free verse? The project blends coding, AI, and creativity, uncovering hidden relationships across art and literature.
When Pictures Teach: Exploring How AI Connects Science Diagrams and Explanations
Science learning often involves both visuals and text—think of a biology textbook with diagrams of cells or physics books with force diagrams. This project uses a multimodal AI model that embeds both pictures and words in the same space, letting students explore how well scientific diagrams align with short explanations. Does the AI correctly connect a sketch of mitosis with the sentence “cells dividing into two identical copies”? Or does it mix them up with unrelated processes? Students can analyze successes and mistakes to learn about both AI and science education.