Selected work

Representative Papers

Stealthy Poisoning Attacks Bypass Defenses in Regression Settings

Jan 29, 2026

This work addresses the vulnerability of existing regression models to stealthy data poisoning attacks, which often evade detection under unrealistic threat assumptions prevalent in prior research. We present the first systematic characterization of covert poisoning attacks across varying detectability levels and introduce a target-normalization-based evaluation framework to rigorously quantify the trade-off between attack efficacy and detectability, enabling the construction of optimal stealthy attacks. To counter this threat, we propose BayesClean, a novel Bayesian defense mechanism that integrates robust regression with Bayesian inference. Extensive experiments demonstrate that BayesClean significantly outperforms state-of-the-art defenses under high-stealth, large-scale poisoning scenarios and effectively mitigates attacks that circumvent current defense strategies.

0 citationsRead paper

Conflict Detection in AI-RAN: Efficient Interaction Learning and Autonomous Graph Reconstruction

Jan 19, 2026

This work addresses the performance degradation and service disruptions in AI-native mobile networks caused by conflicting objectives among multiple agents. To this end, we propose the first conflict detection framework tailored for AI-RAN, which employs a two-tower encoder to learn agent interactions directly from RAN data. A data-driven sparsification mechanism is introduced to automatically reconstruct conflict graphs without requiring manual thresholds or hyperparameter tuning. By eliminating reliance on complex graph neural networks and handcrafted rules, our approach significantly reduces computational overhead while enhancing the accuracy and adaptability of graph construction. This provides an efficient and robust foundation for cooperative intelligence in 6G AI-native networks.

0 citationsRead paper

Applying Formal Methods Tools to an Electronic Warfare Codebase (Experience report)

Jan 16, 2026

This study addresses the limited adoption of formal methods in electronic warfare systems, which is hindered by steep learning curves, inconsistent terminology, and challenges in integrating with existing development workflows. From the perspective of frontline engineers, the authors empirically evaluate multiple static analysis tools—covering input/output contracts, memory effect analysis, and loop invariants—on safety-critical electronic warfare software, comparing their efficacy against conventional unit testing. The research highlights fundamental differences in reasoning paradigms and tool usage between formal methods and traditional testing, revealing the unique capability of formal techniques to uncover deep-seated security vulnerabilities. Furthermore, the paper identifies concrete directions for improving usability, including reducing manual intervention and enhancing support for library code, thereby facilitating broader practical adoption.

0 citationsRead paper

A Generic Machine Learning Framework for Radio Frequency Fingerprinting

Oct 10, 2025

This work addresses the lack of a general-purpose machine learning framework for radio frequency fingerprinting (RFF). We propose the first unified deep learning framework supporting multiple RFF tasks—transmitter-specific identification, transmitter-data association, and RF emitter clustering. Departing from handcrafted feature engineering, our framework learns robust fingerprint representations end-to-end directly from raw in-phase/quadrature (IQ) signals, integrating deep neural networks with feature enhancement mechanisms to achieve cross-device and cross-scenario generalization. Evaluated on real-world RF datasets, it demonstrates significant improvements in identification accuracy and task adaptability across practical applications including satellite-based monitoring, signals intelligence (SIGINT), and counter-drone systems. By overcoming key limitations of conventional methods in flexibility and scalability, our framework establishes a reusable, general-purpose paradigm for engineering deployment of RFF technology.

0 citationsRead paper

TAR: Teacher-Aligned Representations via Contrastive Learning for Quadrupedal Locomotion

Mar 26, 2025

Quadrupedal robot reinforcement learning suffers from teacher-student representation misalignment, covariate shift induced by behavioral cloning, and poor deployment adaptability—leading to weak real-world generalization. Method: We propose a contrastive learning-based teacher-aligned representation framework that, for the first time, incorporates privileged teacher information into self-supervised contrastive learning to achieve representation alignment between teacher and student in simulation. The framework supports online adaptation and continual fine-tuning without privileged signals during deployment. Contribution/Results: By integrating contrastive learning, representation alignment, and behavioral cloning, our method accelerates training by 2× while improving average out-of-distribution (OOD) generalization performance by 40%—surpassing even the full-privilege teacher policy. It significantly enhances real-world deployment capability and overcomes the generalization bottleneck inherent in conventional behavioral cloning approaches.

0 citationsRead paper

Recent publications

Latest Papers

Stealthy Poisoning Attacks Bypass Defenses in Regression Settings

Jan 29, 2026

This work addresses the vulnerability of existing regression models to stealthy data poisoning attacks, which often evade detection under unrealistic threat assumptions prevalent in prior research. We present the first systematic characterization of covert poisoning attacks across varying detectability levels and introduce a target-normalization-based evaluation framework to rigorously quantify the trade-off between attack efficacy and detectability, enabling the construction of optimal stealthy attacks. To counter this threat, we propose BayesClean, a novel Bayesian defense mechanism that integrates robust regression with Bayesian inference. Extensive experiments demonstrate that BayesClean significantly outperforms state-of-the-art defenses under high-stealth, large-scale poisoning scenarios and effectively mitigates attacks that circumvent current defense strategies.

0 citationsRead paper

Conflict Detection in AI-RAN: Efficient Interaction Learning and Autonomous Graph Reconstruction

Jan 19, 2026

This work addresses the performance degradation and service disruptions in AI-native mobile networks caused by conflicting objectives among multiple agents. To this end, we propose the first conflict detection framework tailored for AI-RAN, which employs a two-tower encoder to learn agent interactions directly from RAN data. A data-driven sparsification mechanism is introduced to automatically reconstruct conflict graphs without requiring manual thresholds or hyperparameter tuning. By eliminating reliance on complex graph neural networks and handcrafted rules, our approach significantly reduces computational overhead while enhancing the accuracy and adaptability of graph construction. This provides an efficient and robust foundation for cooperative intelligence in 6G AI-native networks.

0 citationsRead paper

Applying Formal Methods Tools to an Electronic Warfare Codebase (Experience report)

Jan 16, 2026

This study addresses the limited adoption of formal methods in electronic warfare systems, which is hindered by steep learning curves, inconsistent terminology, and challenges in integrating with existing development workflows. From the perspective of frontline engineers, the authors empirically evaluate multiple static analysis tools—covering input/output contracts, memory effect analysis, and loop invariants—on safety-critical electronic warfare software, comparing their efficacy against conventional unit testing. The research highlights fundamental differences in reasoning paradigms and tool usage between formal methods and traditional testing, revealing the unique capability of formal techniques to uncover deep-seated security vulnerabilities. Furthermore, the paper identifies concrete directions for improving usability, including reducing manual intervention and enhancing support for library code, thereby facilitating broader practical adoption.

0 citationsRead paper

A Generic Machine Learning Framework for Radio Frequency Fingerprinting

Oct 10, 2025

This work addresses the lack of a general-purpose machine learning framework for radio frequency fingerprinting (RFF). We propose the first unified deep learning framework supporting multiple RFF tasks—transmitter-specific identification, transmitter-data association, and RF emitter clustering. Departing from handcrafted feature engineering, our framework learns robust fingerprint representations end-to-end directly from raw in-phase/quadrature (IQ) signals, integrating deep neural networks with feature enhancement mechanisms to achieve cross-device and cross-scenario generalization. Evaluated on real-world RF datasets, it demonstrates significant improvements in identification accuracy and task adaptability across practical applications including satellite-based monitoring, signals intelligence (SIGINT), and counter-drone systems. By overcoming key limitations of conventional methods in flexibility and scalability, our framework establishes a reusable, general-purpose paradigm for engineering deployment of RFF technology.

0 citationsRead paper

TAR: Teacher-Aligned Representations via Contrastive Learning for Quadrupedal Locomotion

Mar 26, 2025

Quadrupedal robot reinforcement learning suffers from teacher-student representation misalignment, covariate shift induced by behavioral cloning, and poor deployment adaptability—leading to weak real-world generalization. Method: We propose a contrastive learning-based teacher-aligned representation framework that, for the first time, incorporates privileged teacher information into self-supervised contrastive learning to achieve representation alignment between teacher and student in simulation. The framework supports online adaptation and continual fine-tuning without privileged signals during deployment. Contribution/Results: By integrating contrastive learning, representation alignment, and behavioral cloning, our method accelerates training by 2× while improving average out-of-distribution (OOD) generalization performance by 40%—surpassing even the full-privilege teacher policy. It significantly enhances real-world deployment capability and overcomes the generalization bottleneck inherent in conventional behavioral cloning approaches.

0 citationsRead paper