arXiv Science⌕ Search

arXiv subjects

Shan Du

Publications and source records attributed to Shan Du.

2 recordsLinked to original sources
arXiv2026

FaceLinkGen: A Re-evaluation of Identity Leakage in Privacy-Preserving Face Recognition and Face Anonymization Systems Using Simple Distillation

Privacy-preserving face recognition (PPFR) hides facial appearance while retaining machine recognition, whereas perception-preserving face de-identification (De-ID) retains human recognizability while blocking face recognition models. We show that both designs preserve identity signals that an adaptive attacker can extract. We introduce FaceLinkGen, a unified distillation attack that learns from paired protected and original images. Across MinusFace, PartialFace, and DecoyFace, FaceLinkGen regenerates faces accepted by Face++ at rates of 81.0--99.0\% and by Amazon at rates of 74.9--99.2\%. Against four De-ID methods, it raises mixed-gallery Recall@1 to 55.2--89.6\% and remains effective with limited paired data. These results establish adaptive evaluation as essential for measuring facial privacy and show that resistance to fixed recognition or reconstruction models does not prevent identity leakage.

cs.CV↗
arXiv2026

JVLGS: Joint Vision-Language Gas Leak Segmentation

Gas leaks pose severe risks to human health and industrial safety. However, accurate and timely monitoring of gas leaks remains a major challenge. Existing vision-based methods using infrared (IR) imagery are limited by the inherently blurry and non-rigid nature of leak plumes, which reduces detection reliability and precision. To overcome these limitations, this paper proposes a Joint Vision-Language Gas leak Segmentation (JVLGS) framework that integrates the complementary strengths of visual and textual modalities to enhance gas leak segmentation. Recognizing that gas leaks are sporadic and many video frames contain no leakage, JVLGS incorporates an adaptive postprocessing module to effectively suppress false positives caused by noise and non-target objects-a common limitation of existing approaches. Extensive experiments across diverse industrial scenarios demonstrate that JVLGS significantly outperforms state-of-the-art gas leak segmentation methods. Furthermore, it achieves consistently strong performance under both supervised and few-shot learning settings, whereas competing methods typically perform well in only one setting or underperform in both.

cs.CV↗