Scholarlink

In recent years, Vision Transformers (ViTs) have gained increasing attention in the field of face anti-spoofing. However, most methods use ViT-Base (over 86M parameters), which is unsuitable or inefficient for mobile/embedded deployment. While MobileViT reduces size vs. ViT-Base, its direct application to face anti-spoofing shows suboptimal generalization. To tackle this problem, we adopt MobileViT as a lightweight backbone and propose a novel dual-stream token difference adapter for fine-tuning. Our proposed adapter captures fine-grained information from dual directions, enhancing the model’s adaptation ability. Extensive intra-domain and cross-domain experiments show our method’s effectiveness.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adapting Vision Transformer with Dual Stream Token Difference for Mobile Face Anti-spoofing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Liepiao Zhang,
  • Kun Liu,
  • Junduan Huang,
  • Zitong Yu,
  • Wenxiong Kang

摘要

In recent years, Vision Transformers (ViTs) have gained increasing attention in the field of face anti-spoofing. However, most methods use ViT-Base (over 86M parameters), which is unsuitable or inefficient for mobile/embedded deployment. While MobileViT reduces size vs. ViT-Base, its direct application to face anti-spoofing shows suboptimal generalization. To tackle this problem, we adopt MobileViT as a lightweight backbone and propose a novel dual-stream token difference adapter for fine-tuning. Our proposed adapter captures fine-grained information from dual directions, enhancing the model’s adaptation ability. Extensive intra-domain and cross-domain experiments show our method’s effectiveness.