[Paper Review] Structure Destruction and Content Combination for Face Anti-Spoofing

Ke-Yue Zhang, Taiping Yao+5 more|arXiv (Cornell University)|Jul 22, 2021

Biometric Identification and Security28 references4 citations

TL;DR

This paper proposes a novel Destruction and Combination Network (DCN) for face anti-spoofing that enhances robustness by destroying facial structure and combining content from different subdomains. By using Structure Destruction and Content Combination Modules, along with Local Relation Modeling, the method achieves state-of-the-art performance with an ACER of 0.62% on Oulu-NPU, demonstrating superior generalization and feature discrimination without auxiliary supervision.

ABSTRACT

In pursuit of consolidating the face verification systems, prior face anti-spoofing studies excavate the hidden cues in original images to discriminate real persons and diverse attack types with the assistance of auxiliary supervision. However, limited by the following two inherent disturbances in their training process: 1) Complete facial structure in a single image. 2) Implicit subdomains in the whole dataset, these methods are prone to stick on memorization of the entire training dataset and show sensitivity to nonhomologous domain distribution. In this paper, we propose Structure Destruction Module and Content Combination Module to address these two imitations separately. The former mechanism destroys images into patches to construct a non-structural input, while the latter mechanism recombines patches from different subdomains or classes into a mixup construct. Based on this splitting-and-splicing operation, Local Relation Modeling Module is further proposed to model the second-order relationship between patches. We evaluate our method on extensive public datasets and promising experimental results to demonstrate the reliability of our method against state-of-the-art competitors.

Motivation & Objective

• To address the overreliance on global facial structure in existing face anti-spoofing methods.
• To mitigate domain bias caused by implicit subdomains in training datasets.
• To improve model generalization across diverse attack types and unseen domains.
• To enhance feature discrimination by modeling local patch relationships.

Proposed method

• Structure Destruction Module fragments input images into randomly permuted patches to eliminate global facial structure influence.
• Content Combination Module mixes patches from different classes or subdomains into a single image to promote domain-invariant learning.
• Local Relation Modeling Module uses cosine similarity to model second-order relationships between patches, enhancing feature consistency for the same class.
• The framework jointly optimizes for local detail extraction and cross-patch relationship learning to improve robustness.
• A simplified backbone is used, demonstrating that performance gains come from architectural design, not model complexity.
• The method avoids auxiliary supervision like depth maps or r-ppg signals, relying instead on data recombination and self-supervision.

Experimental results

Research questions

• RQ1 Can destroying facial structure in input images reduce overfitting to global facial patterns in face anti-spoofing?
• RQ2 Does combining patches from different subdomains or classes improve model generalization across domains?
• RQ3 Can modeling second-order relationships between image patches enhance feature discrimination in anti-spoofing?
• RQ4 How does the proposed method compare to state-of-the-art approaches in cross-dataset evaluation?
• RQ5 To what extent does the absence of auxiliary supervision affect performance in the proposed framework?

Key findings

• The proposed method achieves an ACER of 0.62% on Oulu-NPU protocol 1, outperforming all prior methods.
• On the SiW dataset, the method achieves the best performance in protocols 1 and 2, with comparable results in protocol 3.
• The ablation study confirms that all three modules—Structure Destruction, Content Combination, and Local Relation Modeling—contribute progressively to performance improvement.
• Feature visualization shows that the Structure Destruction Module reduces reliance on facial edges and structural cues.
• The Content Combination Module significantly improves feature separation between live and spoof samples across domains.
• The Local Relation Modeling Module enhances intra-class compactness and inter-class separability in the feature space, as confirmed by t-SNE visualizations.