arXiv subjects

Dan Du

Publications and source records attributed to Dan Du.

4 recordsLinked to original sources

arXiv2026

Why Are LLM Backdoor Defenses Fragmented? A Feature-Level Explanation with Sparse Autoencoders

Backdoor attacks pose a serious threat to large language models (LLMs), but existing defenses remain fragmented, failing to pro?vide unified defense against both dirty-label and clean-label attacks. To investigate why such fragmentation arises, we present the first systematic feature-level mechanistic analysis of LLM backdoors using sparse autoencoders (SAEs). Starting from a 2 x 2 comparison of clean and poisoned models on clean and triggered inputs, we trace backdoor-induced logit shifts to high-contributing SAE features and categorize them into four roles: interac?tion, suppressed, mixed, and weight-modified features. This taxonomy reveals system?atic encoding differences: dirty-label back?doors are dominated by isolated interaction features, whereas clean-label backdoors rely more on heterogeneous mixtures of mixed and weight-modified features. These differ?ences explain why existing defenses remain fragmented across attack paradigms. We val?idate this hypothesis through inference-time feature clamping, which reduces ASR to at most 10.8% in most dirty-label settings and at most 15.4% in the majority of clean-label settings, while preserving benign-task perfor?mance. These results show that SAE-based analysis can explain defense fragmentation and guide interpretable backdoor mitigation.

cs.CR↗

arXiv2026

Three Dimensional Multiphysics Modelling of Helicon Wave Heating and Antenna Plasma Coupling for Boundary Density Control in Toroidal Fusion Plasmas

Active control of scrape off layer density is emerging as a critical requirement for improving ion cyclotron resonance heating and enabling high performance steady state operation in future magnetic confinement fusion devices. Helicon wave excitation offers a promising physics based approach to generating high density boundary plasmas with high ionization efficiency and low impurity release. In this work, we develop THEMIS code, a fully three dimensional (3D) multiphysics model of helicon wave propagation and power deposition in a toroidal fusion relevant configuration, employing a finite temperature thermal dielectric tensor. The code quantifies the relative contributions of Doppler shifted cyclotron damping, anomalous Doppler damping, collisional damping, and Landau damping, and demonstrates that slow wave propagation and electron Landau damping dominate the accessible heating regime in Helimak device. A comparative study of four planar antenna geometries under the present protruding window configuration shows that geometric cutoffs and SOL density gradients severely limit power penetration into the core accessible region. To address this constraint, we introduce a recessed window launch scheme that positions the dielectric window inside the vacuum vessel and perform systematic parameter scans of window position, antenna geometry, and installation orientation. From these analyses, we identify the key physics driven principles governing efficient helicon wave coupling: the importance of open circuit termination, maximized strap length and width, controlled inter turn spacing, and sufficient clearance from metallic walls to avoid near field suppression. Guided by these principles, we designed an optimized racetrack spiral antenna that increases coupling efficiency by more than an order of magnitude compared with conventional short circuited rectangular spiral antenna.

physics.plasm-ph↗

arXiv2025

QLPro: Automated Code Vulnerability Discovery via LLM and Static Code Analysis Integration

We introduce QLPro, a vulnerability detection framework that systematically integrates LLMs and static analysis tools to enable comprehensive vulnerability detection across entire open-source projects.We constructed a new dataset, JavaTest, comprising 10 open-source projects from GitHub with 62 confirmed vulnerabilities. CodeQL, a state-of-the-art static analysis tool, detected only 24 of these vulnerabilities while QLPro detected 41. Furthermore, QLPro discovered 6 previously unknown vulnerabilities, 2 of which have been confirmed as 0-days.

cs.SE↗

arXiv2022

Learning High-quality Proposals for Acne Detection

Acne detection is crucial for interpretative diagnosis and precise treatment of skin disease. The arbitrary boundary and small size of acne lesions lead to a significant number of poor-quality proposals in two-stage detection. In this paper, we propose a novel head structure for Region Proposal Network to improve the proposals' quality in two ways. At first, a Spatial Aware Double Head(SADH) structure is proposed to disentangle the representation learning for classification and localization from two different spatial perspectives. The proposed SADH ensures a steeper classification confidence gradient and suppresses the proposals having low intersection-over-union(IoU) with the matched ground truth. Then, we propose a Normalized Wasserstein Distance prediction branch to improve the correlation between the proposals' classification scores and IoUs. In addition, to facilitate further research on acne detection, we construct a new dataset named AcneSCU, with high-resolution imageries, precise annotations, and fine-grained lesion categories. Extensive experiments are conducted on both AcneSCU and the public dataset ACNE04, and the results demonstrate the proposed method could improve the proposals' quality, consistently outperforming state-of-the-art approaches. Code and the collected dataset are available in https://github.com/pingguokiller/acnedetection.

cs.CV↗