Institution profile

VinUniversity

Academic institutionasia · vn
Official website
Research library253linked papers
Opportunities0open roles
Selected work

Representative Papers

Benchmarking Cross-Domain Audio-Visual Deception Detection

May 11, 2024arXiv.org

Current audio-visual spoofing detection methods suffer from poor cross-scenario generalization and lack a standardized cross-domain evaluation benchmark. To address this, we introduce the first unified, standardized cross-domain benchmark for audio-visual spoofing detection, supporting both single-source-to-single-target and multi-source-to-single-target domain adaptation settings. We propose MM-IDGM, a gradient-coordinated optimization algorithm, and Attention-Mixer, a novel multimodal fusion architecture. Additionally, we design three novel multi-source domain sampling strategies and integrate OpenSMILE/ResNet-50 feature extractors with CNN/RNN/Transformer backbones. Extensive experiments demonstrate that our approach achieves an average accuracy improvement of 5.2% under the multi-source-to-single-target setting, significantly enhancing cross-domain generalization. The benchmark and methodology provide a reproducible, comparable, and realistic evaluation framework for practical deployment.

2 citationsRead paper

Defending against Model Inversion Attacks via Random Erasing

Sep 02, 2024arXiv.org

To address privacy leakage from model inversion (MI) attacks, this paper pioneers the adaptation of Random Erasing (RE)—originally a data augmentation technique—into a data-level privacy defense mechanism. During training, RE probabilistically erases random rectangular regions in input images, thereby actively degrading the model’s capacity to encode fine-grained private details. The method requires no architectural modifications or loss-function alterations, ensuring orthogonality and compatibility with existing defenses, and effectively alleviates the privacy–utility trade-off. Evaluated across 23 diverse experimental settings, our approach achieves state-of-the-art privacy–utility balance: reconstructed images suffer a substantial PSNR degradation of 12.6 dB, while classification accuracy drops by less than 1.2%. It consistently outperforms mainstream defense strategies in both privacy preservation and task performance.

1 citationsRead paper
Recent publications

Latest Papers