ABDS: Classifying Protocol Divergence Causes via AS-Boundary Correlation
本文提出ABDS方法，通过关联BGP拓扑自动分类协议分歧原因，并发现安全设备返回localhost签名以隐藏IP，提高了网络问题解决效率。
0 citationsRead paper
Selected work
Representative Papers
PALS: Percentile-Aware Layerwise Sparsity for LLM Pruning
This work addresses the limitation of existing one-shot pruning methods for large language models, which typically apply a uniform sparsity ratio across all layers and ignore inter-layer importance variations. The authors propose an adaptive sparsity allocation strategy based on the 99th percentile of activation magnitudes per layer, enabling layer-aware pruning within ±5% of the target sparsity without fine-tuning and with minimal computational overhead. Their findings challenge the common assumption that gradient magnitude is a reliable indicator for weight removal impact. On LLaMA-2-7B at 50% sparsity, the method achieves a WikiText-2 perplexity of 10.96, significantly outperforming uniform pruning (12.92, p<0.001); however, improvements are marginal on LLaMA-3-8B and Mistral-7B.
0 citationsRead paper
Beyond Gradient-Based Attacks: Adversarial Robustness and Explainability Stability in Cybersecurity Classifiers
This study addresses the dual vulnerability of machine learning models in cybersecurity: adversarial attacks not only degrade prediction accuracy but also destabilize SHAP-based interpretability. Focusing on Random Forest and XGBoost classifiers across four tabular cybersecurity datasets, the authors propose an Explainability Stability Index (ESI) and integrate it with Robustness Index (RI) to form a two-dimensional evaluation framework leveraging black-box attack methods such as TreeSHAP, ZOO, and Square Attack. Their analysis reveals that while ZOO attacks yield deceptively high robustness scores on XGBoost (RI ≈ 0.98), they severely compromise explanation stability (ESI = 0.06–0.16). In contrast, Square Attack more faithfully exposes model fragility (RI ≈ 0.36), demonstrating that predictive robustness and explanation stability constitute distinct and independent dimensions of model reliability.
0 citationsRead paper
The Decomposition Is the Fingerprint: Per-Component Identity for Agent Skills
This study addresses the lack of a stable identity mechanism for AI agent skills that preserves semantic similarity, a limitation exacerbated by conventional hashing methods' inability to support fine-grained comparison. To overcome this, the authors propose a locality-sensitive fingerprinting approach based on multiple SimHash instances, which decomposes each skill into three components—prompt, code, and tool—and embeds them separately to generate a fixed 120-byte tripartite fingerprint. This representation enables accurate recovery of skill family identity, precise attribution of reused components, and reliable detection of independently implemented variants, even after rewriting, renaming, refactoring, or cross-language reimplementation. Experimental results demonstrate an AUC of 0.974 across 4,950 skill pairs, a 77-fold reduction in bit overhead, and exact identification of modified components in all 906 tampered samples.
0 citationsRead paper
Leverage Is Not Reach: A Control-Window Law for Single-Neuron Steering in Language Models
Existing theories struggle to predict when single-neuron interventions can effectively steer language model behavior without inducing output collapse. This work proposes the “normalized control window” framework, which uses alignment between the residual stream and the write direction as a unified control coordinate, establishing the first quantitative law for single-neuron controllability. By integrating forward-pass analysis, contrastive filtering, and gradient-free controller identification, the method reveals that effective control requires the behavioral activation threshold to lie below the collapse threshold, distinguishing apparent leverage from actual reachability. Evaluated on 15 held-out neurons, the framework predicts collapse thresholds with a mean absolute error of 0.14 and successfully identifies genuine controllers, demonstrating that only a subset of neurons exerts substantive, actionable influence over refusal behaviors.
0 citationsRead paper
Recent publications
Latest Papers
ABDS: Classifying Protocol Divergence Causes via AS-Boundary Correlation
本文提出ABDS方法，通过关联BGP拓扑自动分类协议分歧原因，并发现安全设备返回localhost签名以隐藏IP，提高了网络问题解决效率。
0 citationsRead paper
PALS: Percentile-Aware Layerwise Sparsity for LLM Pruning
This work addresses the limitation of existing one-shot pruning methods for large language models, which typically apply a uniform sparsity ratio across all layers and ignore inter-layer importance variations. The authors propose an adaptive sparsity allocation strategy based on the 99th percentile of activation magnitudes per layer, enabling layer-aware pruning within ±5% of the target sparsity without fine-tuning and with minimal computational overhead. Their findings challenge the common assumption that gradient magnitude is a reliable indicator for weight removal impact. On LLaMA-2-7B at 50% sparsity, the method achieves a WikiText-2 perplexity of 10.96, significantly outperforming uniform pruning (12.92, p<0.001); however, improvements are marginal on LLaMA-3-8B and Mistral-7B.
0 citationsRead paper
Beyond Gradient-Based Attacks: Adversarial Robustness and Explainability Stability in Cybersecurity Classifiers
This study addresses the dual vulnerability of machine learning models in cybersecurity: adversarial attacks not only degrade prediction accuracy but also destabilize SHAP-based interpretability. Focusing on Random Forest and XGBoost classifiers across four tabular cybersecurity datasets, the authors propose an Explainability Stability Index (ESI) and integrate it with Robustness Index (RI) to form a two-dimensional evaluation framework leveraging black-box attack methods such as TreeSHAP, ZOO, and Square Attack. Their analysis reveals that while ZOO attacks yield deceptively high robustness scores on XGBoost (RI ≈ 0.98), they severely compromise explanation stability (ESI = 0.06–0.16). In contrast, Square Attack more faithfully exposes model fragility (RI ≈ 0.36), demonstrating that predictive robustness and explanation stability constitute distinct and independent dimensions of model reliability.
0 citationsRead paper
The Decomposition Is the Fingerprint: Per-Component Identity for Agent Skills
This study addresses the lack of a stable identity mechanism for AI agent skills that preserves semantic similarity, a limitation exacerbated by conventional hashing methods' inability to support fine-grained comparison. To overcome this, the authors propose a locality-sensitive fingerprinting approach based on multiple SimHash instances, which decomposes each skill into three components—prompt, code, and tool—and embeds them separately to generate a fixed 120-byte tripartite fingerprint. This representation enables accurate recovery of skill family identity, precise attribution of reused components, and reliable detection of independently implemented variants, even after rewriting, renaming, refactoring, or cross-language reimplementation. Experimental results demonstrate an AUC of 0.974 across 4,950 skill pairs, a 77-fold reduction in bit overhead, and exact identification of modified components in all 906 tampered samples.
0 citationsRead paper
Leverage Is Not Reach: A Control-Window Law for Single-Neuron Steering in Language Models
Existing theories struggle to predict when single-neuron interventions can effectively steer language model behavior without inducing output collapse. This work proposes the “normalized control window” framework, which uses alignment between the residual stream and the write direction as a unified control coordinate, establishing the first quantitative law for single-neuron controllability. By integrating forward-pass analysis, contrastive filtering, and gradient-free controller identification, the method reveals that effective control requires the behavioral activation threshold to lie below the collapse threshold, distinguishing apparent leverage from actual reachability. Evaluated on 15 held-out neurons, the framework predicts collapse thresholds with a mean absolute error of 0.14 and successfully identifies genuine controllers, demonstrating that only a subset of neurons exerts substantive, actionable influence over refusal behaviors.
0 citationsRead paper