[Paper Review] Learning Generalized Spoof Cues for Face Anti-spoofing

Haocheng Feng, Zhibin Hong+6 more|arXiv (Cornell University)|May 8, 2020

Biometric Identification and Security37 references54 citations

TL;DR

Reframes face anti-spoofing as anomaly detection with a residual framework that learns generalized spoof cues via a spoof cue generator and an auxiliary classifier, achieving state-of-the-art generalization across datasets.

ABSTRACT

Many existing face anti-spoofing (FAS) methods focus on modeling the decision boundaries for some predefined spoof types. However, the diversity of the spoof samples including the unknown ones hinders the effective decision boundary modeling and leads to weak generalization capability. In this paper, we reformulate FAS in an anomaly detection perspective and propose a residual-learning framework to learn the discriminative live-spoof differences which are defined as the spoof cues. The proposed framework consists of a spoof cue generator and an auxiliary classifier. The generator minimizes the spoof cues of live samples while imposes no explicit constraint on those of spoof samples to generalize well to unseen attacks. In this way, anomaly detection is implicitly used to guide spoof cue generation, leading to discriminative feature learning. The auxiliary classifier serves as a spoof cue amplifier and makes the spoof cues more discriminative. We conduct extensive experiments and the experimental results show the proposed method consistently outperforms the state-of-the-art methods. The code will be publicly available at https://github.com/vis-var/lgsc-for-fas.

Motivation & Objective

• Reformulate FAS as an anomaly detection problem to handle unseen spoof types.
• Learn discriminative live-spoof differences (spoof cues) as spatial maps.
• Use residual learning to amplify spoof cues and improve generalization.
• Train end-to-end with weak supervision for live samples and implicit supervision for spoofs.
• Demonstrate state-of-the-art performance on standard FAS benchmarks.

Proposed method

• Define spoof cues as a pixel-wise map aligned with the input image, zero for live samples and non-zero for spoofs.
• Propose a spoof cue generator (U-Net) with regression loss on live samples to minimize spoof cues for live data.
• Incorporate a multi-scale triplet loss to promote live-live intra-class compactness and live-spoof inter-class separability.
• Use an auxiliary classifier that takes the overlay of the spoof cue map and the input image to amplify spoof cues and improve discrimination.
• Train with a combined loss L = alpha1 Lr + alpha2 sum Lt + alpha3 La, where Lr is pixel-wise regression on live samples, Lt is multi-scale triplet loss, and La is auxiliary classifier loss.
• During testing, compute spoof score as the mean magnitude of the generated spoof cue map (zero center for live).

Experimental results

Research questions

• RQ1 Can FAS be effectively approached as anomaly detection by learning generalized spoof cues that separate live from spoof beyond predefined attack types?
• RQ2 Do residual learning and auxiliary supervision enhance the discrimination and generalization of spoof cues across diverse datasets and attack modalities?
• RQ3 How does learning live-only regression supervision impact detection of unseen spoofs compared to supervising spoofs explicitly?
• RQ4 What is the impact of input strategy (resized vs patched) and multi-scale representations on learned spoof cues and final performance?

Key findings

• The method achieves state-of-the-art intra-dataset performance on SiW and OULU-NPU across protocols.
• It also performs best in inter-dataset testing (CASIA-MFSD vs Replay-Attack) under HTER.
• Visualizations show spoof cues correspond to medium-specific artifacts and enable clear live-spoof separation.
• Ablations confirm the benefit of residual learning, patch-based input, and joint supervision (Lr, Lt, La) for robust spoof cue learning.
• The approach demonstrates strong generalization to unseen spoof types and realistic variations without relying on depth or temporal data.