🚀 Large Language Model | Cryptography | Blockchain
📍 Beijing Institute of Technology
- 📌 Existing cryptographic algorithm libraries used in privacy-preserving machine learning schemes focus on C++-based implementations. However, the tight coupling of this approach with the underlying system architecture and operating system types makes project deployment and portability challenging.
- 📡 Given the powerful code generation capabilities of large language models, we explore their potential to automatically generate practical GPU-friendly algorithm code from CPU-friendly code. We propose a code evaluation benchmark for Fast-NTT and test the generation capabilities of multiple large models. Additionally, we optimize the NTT algorithm based on LUT and (online-offline) working modes to maximize the advantages of GPU parallel computing.
- 🚀 Manuscript: https://arxiv.org/abs/2502.11110
- 🎮 With the continuous development of quantum computing technology, traditional public-key cryptography based on number-theoretic problems is under threat. Lattice cryptography is at the forefront of international research in post-quantum cryptography. This project aims to bridge the gap between discrete Gaussian sampling theory and lattice-based signature implementations.
- 🔧 I studied discrete Gaussian sampling algorithms, digital signature algorithms, and their integration. The modulus was generalized from perfect powers to arbitrary cases, and the sampling process was optimized by constructing sparse, orthogonal linear transformation matrices. A secure digital signature scheme was designed.
- 🤖 This project is part of a national key research and development program. Based on analyzing the data characteristics of mainstream cryptocurrencies (e.g., Bitcoin, Ethereum), it aims to achieve intelligent detection of malicious activities on public blockchains.
- 🚀 I was responsible for using machine learning or network representation learning to analyze phishing scams on public blockchains, mining key information from transaction networks to identify phishing addresses.
🏅 National College Student Mathematical Modeling Competition
- 🥇 National First Prize，Beijing First Prize
- For high-dimensional data modeling problems, a CLR-based high-dimensional sample comparison model was established to process and model the problem.
🏅 Mathematical Contest in Modeling (MCM)
- 🥇 Finalist Award
- For the problem of predicting user data in the "Wordle" game, a BP neural network model was established using the idea of classification modeling. Interesting features of the dataset were explored from both micro and macro perspectives.
🏅 Huazhong Cup – "Quantitative Investment Problem
- 🥇 First Prize
- For the stock prediction problem, a deep LSTM neural network based on PSO and a dynamic programming model were established to solve the quantitative investment problem.