Selected work

Representative Papers

A Generalized Leakage Interpretation of Alpha-Mutual Information

Jan 14, 2026

This work proposes a unified interpretation of α-mutual information in the context of quantitative information flow and its connection to privacy leakage. By constructing an adversarial generalized decision model and employing Kolmogorov–Nagumo averages together with q-logarithms to characterize the adversary’s gain, the study establishes α-mutual information as a specific instance of generalized g-leakage for the first time. The theoretical link between α-mutual information and generalized g-leakage reveals that the parameter α precisely corresponds to the adversary’s degree of risk aversion. This insight yields a cohesive framework for interpreting privacy leakage through the lens of information-theoretic measures, thereby deepening the understanding of the relationship between information metrics and adversarial behavior.

1 citationsRead paper

Knowledge Distillation under Teacher Misspecification: An Order-Parameter Analysis of the Gap between Teacher Mimicry and Task Performance

Aug 29, 2026

研究通过分析教师模型与真实任务间差距，探讨了在教师模型存在偏差时知识蒸馏的有效性问题。

0 citationsRead paper

Statistical Analysis of Executability and Program Equivalence in Decompilation for IoT Vulnerability Detection

Aug 07, 2026

This study addresses the challenge of semantic fidelity in existing decompilation methods for IoT firmware vulnerability detection, noting that conventional evaluation metrics fail to detect structurally altered code that superficially preserves correct behavior. To this end, the authors propose a novel nine-dimensional decompilation quality assessment framework encompassing structural, behavioral, and semantic aspects. They conduct a statistical analysis of 19,625 decompiled samples from 318 OpenWrt programs generated by five state-of-the-art decompilers. Integrating rule-based techniques with large language models (LLMs), they employ Cohen’s d effect size to quantify inter-group differences and find that recompilable outputs achieve significantly higher overall scores (d = 0.92), with behavioral fidelity (d = 0.96) and structural similarity (d = 0.69) emerging as key predictors. This work establishes a generalizable evaluation framework for black-box decompilation models.

0 citationsRead paper

$r$-deformed $α$-$z$-Rényi relative entropy

Jul 02, 2026

This work proposes a generalized Rényi relative entropy that satisfies fundamental information-theoretic axioms, particularly the data processing inequality, and provides a tighter upper bound for the Tsallis relative entropy. To this end, the authors introduce for the first time the r-logarithm function and construct a three-parameter family of r-deformed α-z-Rényi relative entropies. Through rigorous analysis of quantum divergences, characterization of the parameter space, and comparison via operator inequalities, they systematically establish the conditions under which this quantity constitutes a valid divergence. The study precisely identifies the parameter region where the data processing inequality holds and demonstrates that, in the setting of density operators, the derived upper bound strictly improves upon existing results.

0 citationsRead paper

Closed-Form Steepest Descent Direction toward Flat Minima: Reducing Upper Bounds on the Loss Hessian Eigenspectrum in Neural Networks

Jun 26, 2026

This work addresses the challenge of improving neural network generalization by analyzing optimization dynamics through the lens of data distribution and parameter structure, with a focus on steering training toward flat minima. Leveraging the Wolkowicz–Styan inequality, the authors derive, for the first time, a closed-form expression for the gradient of the largest eigenvalue of the Hessian of the cross-entropy loss—enabling explicit optimization of its spectral upper bound without numerical approximation. By updating parameters along the steepest descent direction defined by this gradient, the method effectively compresses the Hessian eigenvalue spectrum in three-layer networks, thereby avoiding sharp minima and saddle points. This approach guides convergence toward flatter minima that exhibit superior generalization, offering a novel theoretical and algorithmic pathway for understanding and promoting flatness in deep learning optimization.

0 citationsRead paper

Recent publications

Latest Papers

Knowledge Distillation under Teacher Misspecification: An Order-Parameter Analysis of the Gap between Teacher Mimicry and Task Performance

Aug 29, 2026

研究通过分析教师模型与真实任务间差距，探讨了在教师模型存在偏差时知识蒸馏的有效性问题。

0 citationsRead paper

Statistical Analysis of Executability and Program Equivalence in Decompilation for IoT Vulnerability Detection

Aug 07, 2026

This study addresses the challenge of semantic fidelity in existing decompilation methods for IoT firmware vulnerability detection, noting that conventional evaluation metrics fail to detect structurally altered code that superficially preserves correct behavior. To this end, the authors propose a novel nine-dimensional decompilation quality assessment framework encompassing structural, behavioral, and semantic aspects. They conduct a statistical analysis of 19,625 decompiled samples from 318 OpenWrt programs generated by five state-of-the-art decompilers. Integrating rule-based techniques with large language models (LLMs), they employ Cohen’s d effect size to quantify inter-group differences and find that recompilable outputs achieve significantly higher overall scores (d = 0.92), with behavioral fidelity (d = 0.96) and structural similarity (d = 0.69) emerging as key predictors. This work establishes a generalizable evaluation framework for black-box decompilation models.

0 citationsRead paper

$r$-deformed $α$-$z$-Rényi relative entropy

Jul 02, 2026

This work proposes a generalized Rényi relative entropy that satisfies fundamental information-theoretic axioms, particularly the data processing inequality, and provides a tighter upper bound for the Tsallis relative entropy. To this end, the authors introduce for the first time the r-logarithm function and construct a three-parameter family of r-deformed α-z-Rényi relative entropies. Through rigorous analysis of quantum divergences, characterization of the parameter space, and comparison via operator inequalities, they systematically establish the conditions under which this quantity constitutes a valid divergence. The study precisely identifies the parameter region where the data processing inequality holds and demonstrates that, in the setting of density operators, the derived upper bound strictly improves upon existing results.

0 citationsRead paper

Closed-Form Steepest Descent Direction toward Flat Minima: Reducing Upper Bounds on the Loss Hessian Eigenspectrum in Neural Networks

Jun 26, 2026

This work addresses the challenge of improving neural network generalization by analyzing optimization dynamics through the lens of data distribution and parameter structure, with a focus on steering training toward flat minima. Leveraging the Wolkowicz–Styan inequality, the authors derive, for the first time, a closed-form expression for the gradient of the largest eigenvalue of the Hessian of the cross-entropy loss—enabling explicit optimization of its spectral upper bound without numerical approximation. By updating parameters along the steepest descent direction defined by this gradient, the method effectively compresses the Hessian eigenvalue spectrum in three-layer networks, thereby avoiding sharp minima and saddle points. This approach guides convergence toward flatter minima that exhibit superior generalization, offering a novel theoretical and algorithmic pathway for understanding and promoting flatness in deep learning optimization.

0 citationsRead paper

Analytical Evaluation of DCA Convergence Properties for Minimizing Prediction Functions of Gaussian RBF Support Vector Regression

Jun 02, 2026

This study addresses the lack of a priori convergence assessment for the Difference-of-Convex Algorithm (DCA) in Gaussian radial basis function kernel support vector regression (RBF-SVR). By exploiting the analytical structure of the RBF kernel, the authors construct an explicit DC decomposition and, for the first time, derive closed-form expressions for the lower bound μ of the strong convexity parameter and the upper bound L of the gradient Lipschitz constant of the DC components. They further identify the scalar Cαρ—determined by the hyperparameters C and γ—as the key quantity governing DCA’s convergence rate and dependence on initialization. Numerical experiments on six benchmark functions confirm that Cαρ alone effectively predicts DCA’s convergence behavior both before and after training, establishing the first analytical link between RBF-SVR hyperparameters and DCA convergence.

0 citationsRead paper