Research map / Vision-language and multimodal / Visual question answering

Interpretable Face Anti-Spoofing: Enhancing Generalization with Multimodal Large Language Models

AAAI 2025 · Vision-language and multimodal › Visual question answering · approach: Multimodal LLM question answering · 14 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Reformulates face anti-spoofing as an interpretable VQA task using an MLLM (frozen CLIP ViT-L/14 + OPT-2.7B) trained with spoof-aware captions, a lopsided LM loss, and a globally aware connector, evaluated on standard and One-to-Eleven cross-domain protocols.

Problem

Face anti-spoofing methods are binary classifiers with uninterpretable scores that generalize poorly to new domains and unseen spoof types, and existing text-supervised FAS methods rely on limited manual templates.

Approach

Treat FAS as VQA where the answer includes a judgment (Yes/No) plus a natural-language interpretation; generate spoof-aware captions via keyword-filtered captioner fine-tuning, weight the judgment loss more heavily (L-LM loss), and fuse multi-layer CLS tokens into the connector (GAC).

Key results

• Under One-to-Eleven protocol (train on CelebA-Spoof only), average HTER 11.30% and AUC 93.71% versus 18.73% / 87.90% for FLIP.
• Protocol 1 average HTER 1.33% versus 2.98% for CFPL and 3.01% for FLIP.
• Removing the GAC raises average HTER from 11.30% to 14.84%; removing SCF raises it to 14.47%.

Measured effects reported in the paper

Surprising findings

• Using generic (non-spoof-aware) captions as supervision did not improve FAS judgments, indicating caption diversity alone is insufficient.
• The method is not best on several individual target domains (e.g., CASIA-MFSD, HKBU-MARs-V1+, WMCA, HiFiMask), and the w/o L-LM variant is better on some datasets.

Limitations

• Relies on keyword dictionary filtering and a general captioner, so spoof-aware captions may be noisy.
• Weaker on low-level texture cues (moiré, blur) without the GAC, and mixed per-domain results versus baselines.

Experimental setup

Datasets: CASIA-MFSD Replay-Attack MSU-MFSD OULU-NPU SIW Rose-Youtu HKBU-MARs-V1+ WMCA SIW-M-V2 CASIA-SURF-3DMask CelebA-Spoof HiFiMask

Benchmarks: MSU-MFSD CASIA-MFSD Replay-Attack OULU-NPU SIW Rose-Youtu HKBU-MARs-V1+ WMCA SIW-M-V2 CASIA-SURF-3DMask HiFiMask

Metrics: HTER AUC

Baselines: FLIP ViTAF ViT-B ViT-L CFPL DiVT-M UDG-FAS TTDG SSAN PatchNet

Builds on: FLIP BLIP-2 CLIP OPT

Closest papers by meaning

• FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.94)
• Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing (CVPR 2026 · similarity 0.92)
• SLIP: Spoof-Aware One-Class Face Anti-Spoofing with Language Image Pretraining (AAAI 2025 · similarity 0.90)
• From Intuition to Investigation: A Tool-Augmented Reasoning MLLM Framework for Generalizable Face Anti-Spoofing (CVPR 2026 · similarity 0.90)
• mmFAS: Multimodal Face Anti-Spoofing Using Multi-Level Alignment and Switch-Attention Fusion (AAAI 2025 · similarity 0.89)
• PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning (AAAI 2026 · similarity 0.89)