Institution profile

Kyoto University

Academic institutionasia · jp
Official website
Research library567linked papers
Opportunities0open roles
Selected work

Representative Papers

Jailbreaking and Mitigation of Vulnerabilities in Large Language Models

Oct 20, 2024arXiv.org

Existing research on large language models (LLMs) lacks a unified taxonomy for prompt injection and jailbreaking attacks, and insufficiently evaluates defenses under dynamic, interactive scenarios. Method: We propose the first four-dimensional attack taxonomy—spanning prompt-level, model-level, multimodal, and multilingual pathways—and develop a robust alignment framework tailored to interactive settings, alongside a novel automated jailbreaking detection method. We further conduct systematic defense benchmarking, bias diagnosis of existing evaluation benchmarks, and multi-dimensional security measurement. Contribution/Results: Our analysis reveals critical failure modes of current defenses in dynamic interactions, identifies key research gaps—including ethical implications and data bias—and delivers the first comprehensive technical roadmap for LLM safety alignment.

7 citationsRead paper

The"double"square-root law: Evidence for the mechanical origin of market impact using Tokyo Stock Exchange data

Feb 22, 2025

This paper addresses the long-standing debate on the microfoundations of price impact: whether it arises mechanically from order flow or informationally from informed trading. Using high-frequency, trader-identified order-level data from the Tokyo Stock Exchange (2012–2018), we provide the first empirical evidence of the square-root impact law at the individual order level and discover that its temporal decay follows an inverse square-root pattern—collectively termed the “double square-root law”: impact ∝ √volume × 1/√time. Through meta-order reconstruction, anonymized control experiments, and nonparametric impact curve estimation, we demonstrate the robustness of this law and show that synthetically reconstructed meta-orders replicate observed impact dynamics. Our findings strongly support a purely mechanical origin of price impact, offering the first high-resolution empirical validation for market microstructure theory and challenging the dominant informational paradigm.

2 citationsRead paper

Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning

Feb 07, 2024

This work investigates how dataset characteristics affect vulnerability to membership inference attacks (MIAs) in deep transfer learning, particularly for non-differentially private (non-DP) fine-tuned models. Method: We systematically quantify the impact of dataset size and per-class sample count on MIA success rates through empirical evaluation and theoretical modeling based on a simplified fine-tuning process. Contribution/Results: We establish, for the first time, that MIA advantage decays as a power law with respect to per-class sample count under fixed false positive rate—a finding that reveals the impractically large sample requirements needed to robustly protect the most vulnerable samples. This bridges a critical gap between DP-based privacy theory and real-world MIA threat models. Our results yield actionable, quantifiable guidelines for dataset-scale design in transfer learning, significantly enhancing the interpretability and controllability of privacy risks in non-DP settings.

2 citationsRead paper

Hardware-Friendly Delayed-Feedback Reservoir for Multivariate Time-Series Classification

Nov 01, 2022IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems

In edge computing for multivariate time-series classification, conventional reservoir computing (RC) suffers from variable input lengths, resulting in non-fixed feature dimensions; existing intermediate representation (IR) methods rely on computationally expensive matrix inversion, hindering hardware deployment. Method: This paper proposes a dot-product-based fixed-dimensional intermediate representation (DPRR) and a fully digital delay-feedback reservoir (DFR), establishing the first purely digital RC architecture—eliminating analog circuitry and enabling high-level synthesis (HLS). DPRR achieves low-complexity, fixed-dimension feature compression via dot-product mapping, while DFR replaces analog dynamics with digital delay loops to drastically reduce hardware resource consumption. Contribution/Results: FPGA implementations across 12 standard benchmarks demonstrate superior classification accuracy over state-of-the-art methods, alongside 37–62% reduction in logic utilization—achieving a synergistic breakthrough in both accuracy and hardware efficiency.

2 citationsRead paper

Graph Burning: Bounds and Hardness

Feb 29, 2024arXiv.org

This paper investigates the computational complexity and theoretical bounds of the graph burning number: given a graph $G$, one unburnt vertex is ignited per step, and its neighbors burn automatically in the next step; the goal is to minimize the number of steps required to burn the entire graph. Methodologically, the authors employ combinatorial graph theory, structural analysis of graph classes, and carefully constructed polynomial-time reductions. Their contributions include: (i) the first proof that graph burning remains NP-complete on connected cubic graphs and connected proper interval graphs; (ii) a tight additive-1 upper bound on the burning number for connected $P_k$-free graphs; and (iii) a systematic complexity classification of two natural variants—edge burning (igniting edges only) and total burning (igniting vertices or edges)—establishing their equivalence to the original vertex-burning problem. These results strengthen support for the conjecture that the burning number is at most $lceil sqrt{n} ceil$, and yield tight theoretical bounds for multiple fundamental graph classes while fully mapping the complexity landscape of the variants.

1 citations1 influentialRead paper
Recent publications

Latest Papers