PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Meta Model for Zero- and Few-shot Face Anti-spoofing

Yunxiao Qin, Chenxu Zhao, Xiangyu Zhu, Zezheng Wang, Zitong Yu, Tianyu Fu, Feng Zhou, Jingping Shi, Zhen Lei

arXiv:1904.12490v3cs.CV

TL;DR

Face anti-spoofing methods trained on predefined attacks face the challenge of unseen attacks and costly labeled-data collection for new ones. AIM-FAS applies meta-learning to zero- and few-shot FAS and is reported to outperform existing methods on the proposed benchmarks and existing zero-shot protocols.

  • Problem

    Supervised FAS detectors trained on predefined attacks need extensive data, while new attacks and costly labeled-data collection complicate adaptation.

    Source: §1 · p. 1

  • Method

    AIM-FAS meta-trains a learner to detect unseen attacks from predefined faces and adapt using a few examples of new attacks.

    Source: §1 · p. 2, §3.1 · p. 3, §3.2 · p. 4

  • Results

    AIM-FAS outperforms existing methods on the proposed zero- and few-shot benchmarks and existing zero-shot protocols.

    Source: §1 · p. 2, §6 · p. 7

  • Takeaways & Limitations

    The paper frames FAS as a simultaneous zero- and few-shot learning problem and proposes three benchmarks to assess it.

    Source: §1 · p. 2

    On CASIA, limited variation in Cut Photo and Warped Photo categories can make it difficult for the meta-learner to capture discrimination effective for the Video category.

    Source: §5.3 · p. 6

Abstract

from arXiv · show

Face anti-spoofing is crucial to the security of face recognition systems. Most previous methods formulate face anti-spoofing as a supervised learning problem to detect various predefined presentation attacks, which need large scale training data to cover as many attacks as possible. However, the trained model is easy to overfit several common attacks and is still vulnerable to unseen attacks. To overcome this challenge, the detector should: 1) learn discriminative features that can generalize to unseen spoofing types from predefined presentation attacks; 2) quickly adapt to new spoofing types by learning from both the predefined attacks and a few examples of the new spoofing types. Therefore, we define face anti-spoofing as a zero- and few-shot learning problem. In this paper, we propose a novel Adaptive Inner-update Meta Face Anti-Spoofing (AIM-FAS) method to tackle this problem through meta-learning. Specifically, AIM-FAS trains a meta-learner focusing on the task of detecting unseen spoofing types by learning from predefined living and spoofing faces and a few examples of new attacks. To assess the proposed approach, we propose several benchmarks for zero- and few-shot FAS. Experiments show its superior performances on the presented benchmarks to existing methods in existing zero-shot FAS protocols.

1 Introduction

Conventional face anti-spoofing detectors are trained on predefined attacks, while new attacks and costly data collection leave deployment challenges. The paper frames FAS as a zero- and few-shot problem and introduces AIM-FAS and three benchmarks to address it.

1 Introduction

  • New, evolving presentation attacks and out-of-distribution living faces can produce unpredictable results, while adapting to new attacks requires costly labeled data.Existing methods need sufficient samples to train for each new attack.Source: §1 · p. 1
  • The paper formulates FAS as zero- and few-shot learning: generalize from predefined attacks to unseen attacks, then adapt using a few examples of a new attack.Few-shot learning uses both predefined examples and the collected examples of the new attack.Source: §1 · pp. 1–2
  • The paper proposes three benchmarks for assessing detection of new presentation attacks from the same domain, different domains, and different modalities.Source: §1 · p. 2
  • AIM-FAS uses meta-learning and Fusion Training to focus on both detecting unseen attacks without examples and adapting when a few new-attack examples are available.Its Adaptive Inner-Update strategy uses a learnable inner-update step size.Source: §1 · p. 2
  • AIM-FAS achieves state-of-the-art results on the proposed zero- and few-shot anti-spoofing benchmarks.The paper also evaluates the method on existing zero-shot protocols.Source: §1 · p. 2

2 Background

The background reviews conventional face anti-spoofing and the learning approaches relevant to the paper’s zero- and few-shot framing.

2.1 Face Anti-Spoofing

  • Traditional and deep-learning FAS methods distinguish live from spoof faces, but remain sensitive to illumination, blur, capture camera, and attack instruments.Depth-regression methods are reported to outperform binary-classification methods, mainly because they provide more detail about spoofing cues.Source: §2.1 · p. 2

2.2 Few-shot and Zero-shot Learning

  • Few-shot meta-learning trains on tasks so a learner can recognize unseen categories from a support set of few labeled examples and a query set.Source: §2.2 · p. 2
  • Traditional zero-shot learning recognizes unseen categories using descriptions or semantic attributes, while this paper addresses zero- and few-shot FAS together.Source: §2.2 · p. 2

3 Methodology

AIM-FAS trains a meta-learner on zero- and few-shot face anti-spoofing tasks, with support-set adaptation and query-set evaluation. It combines adaptive inner updates with training across different shot counts.

3.1 Zero- and Few-shot FAS Problem

  • Zero-shot FAS predicts new living and spoofing categories from predefined categories; few-shot FAS also uses a few examples of new categories.The zero-shot support set contains predefined faces, while the few-shot support set includes both predefined and new living and spoofing faces.Source: §3.1 · p. 3

Task generation

  • Tasks pair living and spoofing categories, using predefined categories for support and new categories for queries; K determines whether new-category examples also enter support.The figure describes generating zero- and few-shot tasks from fine-grained living and spoofing categories, then adapting on support and evaluating on query.Source: Task generation · p. 3, Figure 2 · p. 3

3.2 AIM-FAS

  • AIM-FAS adapts its weights on each task’s support set, then uses query-set performance to optimize the meta-learner and its adaptive update parameters.The inner-update learning rate changes with the update step, and query loss updates the meta-learner’s parameters.Source: §3.2 · pp. 3–4

Fusion Train (FT)

  • Fusion Training trains one meta-learner on tasks with different shot counts, including zero-shot tasks, rather than training separate learners for each count.The authors report that Fusion Training improves AIM-FAS on both zero- and few-shot FAS tasks.Source: Fusion Train (FT) · p. 4

Network

  • FAS-DR predicts facial depth from RGB images by concatenating features from three cascaded backbone blocks.The input is a 256×256×3 RGB image, and the predicted depth map is 32×32×1.Source: Network · p. 4, Figure 3 · p. 4

4 Zero- and Few-shot FAS Benchmarks

The paper introduces three zero- and few-shot FAS benchmarks spanning single-domain, cross-domain, and cross-modal evaluation.

OULU-ZF

  • OULU-ZF reorganizes OULU-NPU as a single-domain benchmark, with no overlap between seen training categories and unseen test categories.The test set uses living faces from the living 1 category; training includes other living categories and print and replay attacks.Source: §4 · p. 5

Cross-ZF

  • Cross-ZF tests cross-domain generalization using separate source and target datasets, with 3D-mask attacks in the test set unlike the printed and replayed training attacks.Training draws on CASIA-MFSD, MSU-MFSD, and SiW; testing draws on 3DMAD, Oulu-NPU, and Replay-Attack.Source: §4 · p. 5

SURF-ZF

  • SURF-ZF evaluates cross-modal adaptation, training on RGB and depth and validating and testing on IR and depth modalities.Its stated purpose is to test fast learning from new modalities.Source: §4 · p. 5

5 Experiments

The experiments evaluate AIM-FAS across three proposed zero- and few-shot benchmarks and an existing zero-shot protocol, then examine its components and feature adaptation. AIM-FAS outperforms other detectors across the proposed benchmarks, while results on the existing protocol vary by sub-protocol.

5.1 Experiment Setup

  • The evaluation trains on tasks generated from training data, tests on zero- or few-shot tasks from test data, and calculates performance across test tasks.The proposed benchmarks are OULU-ZF, Cross-ZF, and SURF-ZF; implementation generates depth labels normalized to [0, 1] for living faces and all-zero labels for spoofing faces.Source: §5.1 · p. 5, Table 1 · p. 5, Table 2 · p. 5

5.3 Experiment on Existing Dataset

  • On the existing zero-shot protocol, AIM-FAS improves average AUC by at least 0.52% but is not best on three sub-protocols.The authors suggest unitary training spoof categories may explain weaker performance on CASIA Video, Replay-Attack Video, and MSU Printed Photo.Source: §5.3 · p. 6, Table 6 · p. 7

5.2 Experiment on Proposed Benchmarks

  • AIM-FAS outperforms other detectors on all three proposed benchmarks, with ACER reductions over FAS-DR across zero-, 1-, and 5-shot tasks.On OULU-ZF, reductions are 25%, 17%, and 28%; on Cross-ZF, 38%, 30%, and 38%; on SURF-ZF, 12%, 13%, and 16%, respectively.Source: Table 5 · p. 6, §5.2 · p. 6

5.4 Ablation Study

  • Removing predefined living and spoofing faces from the support set raises ACER by 11% and 42% on 1-shot and 5-shot tasks, respectively.AIM-FAS (Resnet) also lowers ACER versus Resnet-10 by 45.08%, 54.72%, and 41.55% on 0-, 1-, and 5-shot Cross-ZF tasks.Source: §5.4 · p. 6, Table 7 · p. 7
  • Without Fusion Training, separate meta-learners trained for each shot setting perform worse across shot scenes; Adaptive Inner-Update lowers ACER by more than 3.0.The reported ACER reduction applies to 0-, 1-, and 5-shot tasks; learned α and γ rise as ACER falls during meta-training.Source: §5.4 · p. 7, Figure 5 · p. 7, Table 7 · p. 7

5.5 Visualization and Analysis

  • After 50 support-set inner-update steps on a 5-shot task, query-set features become more separated, and AIU helps learn better discrimination.Across the visualized progression, category inner distance L1 declines while inter distance L2 rises.Source: §5.5 · p. 7, Figure 6 · p. 7

6 Conclusion and Future Work

AIM-FAS reframes face anti-spoofing as a simultaneous zero- and few-shot learning problem and outperforms existing methods with a clear margin on both. The paper also proposes three benchmarks and identifies more challenging, practical application scenes as future work.

6 Conclusion and Future Work

  • Experiments show AIM-FAS outperforms existing methods with a clear margin on both zero- and few-shot FAS.Experiments were conducted on the proposed benchmarks and existing zero-shot protocols.Source: §6 · p. 7
  • The paper reframes face anti-spoofing as a simultaneous zero- and few-shot learning issue, develops AIM-FAS, and proposes three benchmarks.Experiments validate AIM-FAS on the proposed benchmarks and existing zero-shot protocols.Source: §6 · p. 7
  • Future work will develop AIM-FAS for more challenging and practical application scenes.Source: §6 · p. 7

Related papers

Loading 1904.12490v3…