2026.08.21
Associate Professor Yuntao Wang (Department of Informatics) and his research team received the Best Paper Award at the 9th International Conference on Frontiers in Cyber Security (FCS 2026), held in Macao from August 3 to 5, 2026.
FCS is an international conference that brings together researchers to present and discuss the latest advances in cybersecurity. The conference covers a broad range of topics, including cryptography, system and network security, and security technologies based on artificial intelligence and machine learning.
The award-winning paper investigates machine-learning-based attacks on the Learning With Errors (LWE) problem, which serves as an important security foundation for post-quantum cryptography (PQC). In recent years, new cryptanalytic approaches using machine-learning techniques such as Transformers have attracted considerable attention for their potential to recover secret information from LWE instances. However, existing approaches require computationally expensive preprocessing based on BKZ lattice basis reduction to generate suitable training data.
In this work, the researchers proposed a low-cost preprocessing approach that analyzes the coefficient distribution induced by BKZ reduction and generates training data that mimic this distribution without repeatedly performing costly BKZ reduction. Experimental results demonstrate that appropriately controlling the coefficient distribution enables the recovery of secrets with higher Hamming weights while substantially reducing the amount of training data required for successful secret recovery. In particular, under the strongest preprocessing condition evaluated, the maximum recoverable Hamming weight increased from 3 to 15. For dimensions 40 and 50, the number of training samples required for secret recovery was reduced to approximately 5-7% of that required when using uniformly distributed inputs.
These results provide a new perspective on evaluating the security of the LWE problem through machine-learning-based cryptanalysis and are expected to contribute to the security evaluation of lattice-based and post-quantum cryptographic schemes.
【Award winner】Yuntao Wang, Kazuma Kubota, Haibo Zhang
【Tittle】Low-Cost Distribution-Mimicking Preprocessing for Machine-Learning Attacks on Binary LWE
【Authors】Yuntao Wang, Kazuma Kubota, Haibo Zhang
