Selected work

Representative Papers

ASSEMBLAGE-DEEPHISTORY: A Cross-Build Binary Dataset with Temporal Coverage

May 20, 2026

Existing binary corpora lack unified modeling of cross-compiler builds, multi-version evolution, and vulnerability labels. This work introduces a queryable binary dataset encompassing 248 open-source projects, multiple compilers and optimization levels, and historical versions spanning several years, integrating build diversity, temporal dynamics, and CVE annotations into a cohesive framework for the first time. The dataset links binaries to their corresponding source code, functions, debug information, and version metadata via database indexing, enabling multidimensional analysis through LLM benchmarks, embedding models (jTrans, MalConv), and TLSH hashing. Experiments reveal that large language models rely on build artifacts rather than semantic reasoning, quantify the clustering behavior of versions in embedding space, and employ Bayesian regression to disentangle the sources of binary similarity, thereby demonstrating the dataset’s value for fine-grained, traceable binary analysis.

0 citationsRead paper

Adversarial Samples Are Not Created Equal

Jan 02, 2026arXiv.org

Existing theoretical frameworks fail to distinguish whether adversarial examples exploit fragile yet predictable non-robust features in data, leading to biased robustness evaluations. This work addresses this gap by formally categorizing adversarial examples into two types: those that rely on non-robust features and those that do not. The authors propose a novel ensemble-based metric to quantify the extent to which adversarial perturbations manipulate non-robust features. By integrating adversarial attack generation with robustness analysis, the proposed framework elucidates the mechanism through which sharpness-aware minimization enhances model robustness and explains the performance discrepancy between standard and adversarial training on robust datasets. This approach offers a refined perspective for evaluating and understanding model robustness.

0 citationsRead paper

Intermediate N-Gramming: Deterministic and Fast N-Grams For Large N and Large Datasets

Nov 18, 2025

Extracting high-frequency n-grams—especially for large n—from massive datasets poses significant challenges in terms of accuracy, efficiency, and determinism. Method: This paper proposes Intergrams, a hardware-aware multi-pass algorithm that exploits the power-law distribution of n-gram frequencies. It generates candidate n-grams from frequent (n−1)-grams, applies frequency-based pruning, and incorporates low-level optimizations to progressively shrink the search space across iterative passes. Theoretical analysis guides algorithm design to ensure exactness and strong scalability. Results: On real-world large-scale datasets, Intergrams achieves 10.3×–33× speedup over the state-of-the-art method. It is the first deterministic approach to break the performance bottleneck for extracting high-frequency n-grams with large n, while guaranteeing correctness and scalability.

0 citationsRead paper

Zipf-Gramming: Scaling Byte N-Grams Up to Production Sized Malware Corpora

Nov 17, 2025

To address the inefficiency of extracting high-frequency 6–8 byte n-grams from terabyte-scale executables—hindering frequent retraining of malware detection models—this paper proposes Zipf-Gramming, a novel n-gram extraction algorithm. It is the first to deeply integrate Zipf’s law into byte-level n-gram engineering, leveraging frequency estimation and approximate sorting to skip low-frequency n-gram computations, thereby eliminating substantial redundant overhead. Evaluated on real-world production data, Zipf-Gramming achieves up to 35× faster top-k high-frequency n-gram extraction compared to state-of-the-art methods, supporting GB/s throughput and sub-10 ms latency. Models built upon these features remain under 2 MB in size while improving AUC by up to 30%. This work establishes a deployable, high-throughput feature extraction paradigm for large-scale, rapidly updated binary malware detection.

0 citationsRead paper

Adversarial Machine Learning Attacks on Financial Reporting via Maximum Violated Multi-Objective Attack

Jul 07, 2025

Financially distressed firms possess strong incentives to manipulate financial statements—concealing operational deterioration, inflating profits, and evading fraud detection—yet existing adversarial attack methods fail to jointly optimize conflicting reverse-correlation objectives (e.g., increasing reported profit while decreasing fraud risk scores). This paper proposes Maximum Violation Multi-Objective (MVMO) attack, a novel adversarial framework that integrates accounting standards and legal constraints to define a realistic threat model, and employs dynamic multi-objective optimization to simultaneously perturb financial metrics and fraud risk scores. Experiments demonstrate that MVMO successfully inflates reported earnings by 100–200% while reducing fraud scores by 15% in ~50% of cases, achieving a 20× higher success rate than conventional methods. To our knowledge, this is the first work to formalize financial statement manipulation as an adversarial machine learning problem, offering novel insights for audit defense mechanisms and regulatory technology (RegTech).

0 citationsRead paper

Recent publications

Latest Papers

ASSEMBLAGE-DEEPHISTORY: A Cross-Build Binary Dataset with Temporal Coverage

May 20, 2026

Existing binary corpora lack unified modeling of cross-compiler builds, multi-version evolution, and vulnerability labels. This work introduces a queryable binary dataset encompassing 248 open-source projects, multiple compilers and optimization levels, and historical versions spanning several years, integrating build diversity, temporal dynamics, and CVE annotations into a cohesive framework for the first time. The dataset links binaries to their corresponding source code, functions, debug information, and version metadata via database indexing, enabling multidimensional analysis through LLM benchmarks, embedding models (jTrans, MalConv), and TLSH hashing. Experiments reveal that large language models rely on build artifacts rather than semantic reasoning, quantify the clustering behavior of versions in embedding space, and employ Bayesian regression to disentangle the sources of binary similarity, thereby demonstrating the dataset’s value for fine-grained, traceable binary analysis.

0 citationsRead paper

Adversarial Samples Are Not Created Equal

Jan 02, 2026arXiv.org

Existing theoretical frameworks fail to distinguish whether adversarial examples exploit fragile yet predictable non-robust features in data, leading to biased robustness evaluations. This work addresses this gap by formally categorizing adversarial examples into two types: those that rely on non-robust features and those that do not. The authors propose a novel ensemble-based metric to quantify the extent to which adversarial perturbations manipulate non-robust features. By integrating adversarial attack generation with robustness analysis, the proposed framework elucidates the mechanism through which sharpness-aware minimization enhances model robustness and explains the performance discrepancy between standard and adversarial training on robust datasets. This approach offers a refined perspective for evaluating and understanding model robustness.

0 citationsRead paper

Intermediate N-Gramming: Deterministic and Fast N-Grams For Large N and Large Datasets

Nov 18, 2025

Extracting high-frequency n-grams—especially for large n—from massive datasets poses significant challenges in terms of accuracy, efficiency, and determinism. Method: This paper proposes Intergrams, a hardware-aware multi-pass algorithm that exploits the power-law distribution of n-gram frequencies. It generates candidate n-grams from frequent (n−1)-grams, applies frequency-based pruning, and incorporates low-level optimizations to progressively shrink the search space across iterative passes. Theoretical analysis guides algorithm design to ensure exactness and strong scalability. Results: On real-world large-scale datasets, Intergrams achieves 10.3×–33× speedup over the state-of-the-art method. It is the first deterministic approach to break the performance bottleneck for extracting high-frequency n-grams with large n, while guaranteeing correctness and scalability.

0 citationsRead paper

Zipf-Gramming: Scaling Byte N-Grams Up to Production Sized Malware Corpora

Nov 17, 2025

To address the inefficiency of extracting high-frequency 6–8 byte n-grams from terabyte-scale executables—hindering frequent retraining of malware detection models—this paper proposes Zipf-Gramming, a novel n-gram extraction algorithm. It is the first to deeply integrate Zipf’s law into byte-level n-gram engineering, leveraging frequency estimation and approximate sorting to skip low-frequency n-gram computations, thereby eliminating substantial redundant overhead. Evaluated on real-world production data, Zipf-Gramming achieves up to 35× faster top-k high-frequency n-gram extraction compared to state-of-the-art methods, supporting GB/s throughput and sub-10 ms latency. Models built upon these features remain under 2 MB in size while improving AUC by up to 30%. This work establishes a deployable, high-throughput feature extraction paradigm for large-scale, rapidly updated binary malware detection.

0 citationsRead paper

Adversarial Machine Learning Attacks on Financial Reporting via Maximum Violated Multi-Objective Attack

Jul 07, 2025

Financially distressed firms possess strong incentives to manipulate financial statements—concealing operational deterioration, inflating profits, and evading fraud detection—yet existing adversarial attack methods fail to jointly optimize conflicting reverse-correlation objectives (e.g., increasing reported profit while decreasing fraud risk scores). This paper proposes Maximum Violation Multi-Objective (MVMO) attack, a novel adversarial framework that integrates accounting standards and legal constraints to define a realistic threat model, and employs dynamic multi-objective optimization to simultaneously perturb financial metrics and fraud risk scores. Experiments demonstrate that MVMO successfully inflates reported earnings by 100–200% while reducing fraud scores by 15% in ~50% of cases, achieving a 20× higher success rate than conventional methods. To our knowledge, this is the first work to formalize financial statement manipulation as an adversarial machine learning problem, offering novel insights for audit defense mechanisms and regulatory technology (RegTech).

0 citationsRead paper