This paper proposes consistent low-rank dequantization as a preprocessing method for gradient inversion attacks under low-bit quantization. It alternates low-rank approximation and projection onto quantization intervals to improve feature recovery from quantized model updates.
@misc{cho2027consistent,title={Consistent Low-Rank Dequantization for Gradient Inversion under Low-Bit Quantization},author={Cho, Seoklea and Gong, Jinu},year={2027},note={Submitted to ICC 2027},}
2025
IEIE
Evaluating Generalization and Data Leakage in Large Language Models via Data Perturbation
This study addresses the issue of data leakage in Large Language Model evaluation and proposes a perturbation-based robustness test to assess true generalization ability. Semantically preserved but surface-level altered datasets are constructed through synonym substitution and label/schema renaming, then evaluated across recent small language models and standard benchmarks.
@misc{kim2025llmleakage,title={Evaluating Generalization and Data Leakage in Large Language Models via Data Perturbation},author={Kim, Taemin and Cho, Seoklea},year={2025},note={IEIE Conference, 2025},}