PaperlayerView exact revision on arXiv

Source-linked AI summary

Dual-Cross Central Difference Network for Face Anti-Spoofing

Zitong Yu, Yunxiao Qin, Hengshuang Zhao, Xiaobai Li, Guoying Zhao

arXiv:2105.01290v1cs.CV

TL;DR

Face anti-spoofing helps secure face recognition, but CDC aggregates gradients from all neighbors and FAS datasets have limited scale and diversity. The paper proposes cross-directional convolutions, a dual-stream network, and Patch Exchange augmentation; it reports state-of-the-art performance across four benchmark datasets and three testing protocols.

  • Problem

    CDC aggregates central gradients from all neighbors, while limited FAS data scale and diversity make task-dedicated augmentation a research need.

    Source: §1 · p. 1

  • Method

    The paper combines horizontal/vertical and diagonal C-CDC streams with CFIM and proposes Patch Exchange augmentation for mixed samples and dense labels.

    Source: §1 · pp. 1–2

  • Results

    The method achieves state-of-the-art performance on four benchmark datasets across intra-dataset, cross-dataset, and cross-type testing protocols.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Patch Exchange can be used with DC-CDN and existing FAS methods for performance improvement.

    Source: §1 · p. 2

Abstract

from arXiv · show

Face anti-spoofing (FAS) plays a vital role in securing face recognition systems. Recently, central difference convolution (CDC) has shown its excellent representation capacity for the FAS task via leveraging local gradient features. However, aggregating central difference clues from all neighbors/directions simultaneously makes the CDC redundant and sub-optimized in the training phase. In this paper, we propose two Cross Central Difference Convolutions (C-CDC), which exploit the difference of the center and surround sparse local features from the horizontal/vertical and diagonal directions, respectively. It is interesting to find that, with only five ninth parameters and less computational cost, C-CDC even outperforms the full directional CDC. Based on these two decoupled C-CDC, a powerful Dual-Cross Central Difference Network (DC-CDN) is established with Cross Feature Interaction Modules (CFIM) for mutual relation mining and local detailed representation enhancement. Furthermore, a novel Patch Exchange (PE) augmentation strategy for FAS is proposed via simply exchanging the face patches as well as their dense labels from random samples. Thus, the augmented samples contain richer live/spoof patterns and diverse domain distributions, which benefits the intrinsic and robust feature learning. Comprehensive experiments are performed on four benchmark datasets with three testing protocols to demonstrate our state-of-the-art performance.

1 Introduction

The introduction frames face anti-spoofing as necessary for securing face recognition and identifies limitations in central difference convolution and FAS data augmentation. It presents C-CDC, DC-CDN, and Patch Exchange, reporting state-of-the-art performance across four benchmark datasets and three testing protocols.

1 Introduction

  • CDC aggregates central gradients from all neighbors, making it inefficient and redundant because gradient directions differ.Source: §1 · p. 1
  • Limited FAS training data and modest gains from generic augmentation motivate a task-dedicated augmentation strategy.The passage attributes limited dataset scale and diversity to high collection costs for spoof generation and video recording.Source: §1 · p. 1
  • C-CDC separates central-gradient features into horizontal/vertical or diagonal directions, using five ninths the parameters and less computation than CDC while achieving better FAS performance.Source: §1 · pp. 1–2
  • DC-CDN uses two streams with horizontal/vertical and diagonal C-CDC, with CFIM for mutual neighbor-relation mining and local-detail enhancement.Source: §1 · p. 2
  • Patch Exchange synthesizes mixed samples with diverse attacks and domains by exchanging face patches and dense labels, and can be used with DC-CDN or existing FAS methods.Source: §1 · p. 2
  • The method achieves state-of-the-art performance on four benchmark datasets across intra-dataset, cross-dataset, and cross-type testing protocols.Source: §1 · p. 2

2 Related Work

Related work covers handcrafted and deep-learning approaches to face anti-spoofing, then discusses limited data and convolution operators for fine-grained features. The paper positions CDC as suitable for FAS while seeking to improve its performance and efficiency.

2 Related Work

  • Handcrafted descriptors capture detailed invariant spoof cues but have limited discriminative capacity, while deep methods use binary or dense pixel-wise supervision, including pseudo depth maps in this work.The cited dense-supervision examples include pseudo depth, reflection, texture, and binary maps.Source: §2 · p. 2
  • Limited public FAS data can leave methods trained on predefined scenarios and attacks prone to overfitting and domain shift; prior work also explores zero-shot detection.Source: §2 · p. 2
  • Prior convolution designs embed local relations or modify spatial aggregation scope, including local relation layers [Hu et al., 2019], dilated convolution [Yu and Koltun, 2015], and deformable convolution.Source: §2 · p. 2
  • CDC is presented as suitable for invariant, detailed FAS features, and this paper aims to improve its performance and efficiency.Source: §2 · p. 2

3 Methodology

The methodology starts from central-difference convolution, sparsifies neighbor aggregation into two cross-direction operators, builds a dual-stream network with feature interaction, and adds patch-exchange augmentation.

3.1 Preliminary

  • CDC augments weighted local-neighbor convolution with central-gradient features to capture detailed cues alongside intensity-level semantics.Source: §3.1 · p. 2

3.2 Cross Central Difference Convolution

  • C-CDC addresses potentially redundant full-neighborhood aggregation by sampling five-position horizontal/vertical or diagonal crosses, whose symmetry is intended to support convergence and robust features.Source: §3.2 · p. 3

3.3 Dual-Cross Central Difference Network

  • C-CDN uses five ninths the parameters and two hundredths the computational cost of DepthNet and CDCN; DC-CDN fuses unshared HV/DG streams through multilevel CFIM.CFIM is designed to mine cross-stream relations and enhance local detail.Source: §3.3 · p. 4

3.4 Patch Exchange Augmentation

  • Patch Exchange swaps face patches across samples with corresponding dense labels, mixing domains and live/attack patterns while encouraging detailed spoof features.Source: §3.4 · pp. 4–5

4 Experiments

Experiments evaluate DC-CDN on four datasets and multiple testing settings, with ablations examining C-CDC, CFIM, and Patch Exchange. The reported comparisons show strong results on intra-dataset, cross-type, and cross-dataset tests.

4.1 Datasets and Metrics

  • The experiments use OULU-NPU for intra-testing, CASIA-MFSD and Replay-Attack for cross-testing, and SiW-M for cross-type testing, with metrics selected by protocol.The reported metrics include ACER, HTER, and EER.Source: §4.1 · p. 5

4.2 Implementation Details

  • Training uses batch size 8, Adam, baseline flip, color-jitter, and Cutout augmentations, and up to 800 epochs; testing averages the predicted depth map.The learning rate starts at 1e-4 and halves at epoch 500; models use MSE and contrastive depth loss.Source: §4.2 · p. 5

4.3 Ablation Study

  • C-CDC(HV) and C-CDC(DG) use five ninths the parameters of CDC and improve performance by 0.3% and 0.1% ACER, respectively.On OULU-NPU Protocol 1, the best ACERs are 0.7% for C-CDC(HV) and 0.9% for C-CDC(DG) at θ = 0.8.Source: §4.3 · p. 5
  • CFIM lowers ACER by 0.2% and 0.4% versus the single-stream C-CDN(HV) and C-CDN(DG), while Patch Exchange improves single- and dual-stream models.Patch Exchange also yields -1.2% EER on cross-type testing and -2.5% and -2.2% HTER on cross-dataset testing.Source: §4.3 · p. 6

4.4 Comparison with State of the Arts

  • DC-CDN ranks first on OULU-NPU’s first three protocols (0.4%, 1.3%, 1.9% ACER), has the best SiW-M EER, and reaches 6.0% HTER on protocol CR.On CR, it outperforms the prior state of the art by 9.5%.Source: Table 3 · p. 5, §4.4 · p. 6, Table 4 · p. 6

5 Conclusions

The conclusion presents C-CDC, DC-CDN with CFIM, and Patch Exchange as contributions to face anti-spoofing. Experiments verify the proposed methods, and future work will extend C-CDC to 3D for fine-grained video understanding.

  • The conclusion proposes C-CDC and a dual-stream DC-CDN with CFIM for feature enhancement in face anti-spoofing.Patch Exchange is proposed to enrich training data.Source: §5 · p. 6
  • The authors report extensive experiments to verify the effectiveness of the proposed methods.Source: §5 · p. 7
  • Future work will extend the C-CDC family to a 3D version for fine-grained video understanding tasks.Source: §5 · p. 7

A. Impacts of Sparsity and Neighbor Locations of C-CDC

The appendix compares sparse and asymmetric neighbor sampling with C-CDC on OULU-NPU Protocol 1. Sparser sampling can sharply reduce performance, while symmetric C-CDC is reported as an efficiency–performance tradeoff.

B. Discussion about CFIM

The appendix examines CFIM’s learned α and β values and initialization. It reports distinct weight trends and the best Protocol 1 result with learnable weights initialized to zero.

  • With zero initialization, α_low, β_mid, and β_high tend to positive values, while α_mid, α_high, and β_low tend to negative values.Source: B. Discussion about CFIM · p. 7
  • The sharp decrease in β_low suggests that low-level features from the C-CDC(DG) stream may contribute little in DC-CDN.Source: B. Discussion about CFIM · p. 7
  • Learnable α and β initialized to zero achieve 0.5% ACER on OULU-NPU Protocol 1; small initial values such as [0, 1] also seem suitable.Initialization values influence CFIM particularly when the weights are learnable.Source: B. Discussion about CFIM · p. 7

C. Analysis of PE Augmentation

Patch Exchange improves performance across FAS architectures and testing scenarios, including small-scale training and unseen attack types. With PE, DC-CDN also localizes partial-attack regions that the model without PE fails to capture.

C. Analysis of PE Augmentation

  • PE consistently improves DepthNet and CDCN, including small-scale OULU-NPU Protocol 4 and unseen attack types in SiW-M.Figure 8 compares PE augmentation across architectures and testing protocols.Source: C. Analysis of PE Augmentation · p. 7, Figure 8 · p. 8
  • With PE, DC-CDN clearly discovers and localizes spoofing regions in synthesized partial-attack faces, unlike the model trained without PE.Both models perform well on normal live faces and print attacks in the first three visualization columns.Source: C. Analysis of PE Augmentation · p. 7
  • Figure 7 compares learnable and fixed CFIM α and β settings with different initial values on OULU-NPU Protocol 1.Source: Figure 7 · p. 8
Loading 2105.01290v1…