Joint Face Spoofing Forgery Eval

joint-face-spoofing-forgery-eval

Benchmarking Joint Face Spoofing and Forgery Detection with Visual and Physiological Cues — Yu et al. (2022) (arXiv:2208.05401, 2022)

What this evaluates

Evaluates models' ability to detect physical face spoofing attacks and digital face forgeries using visual appearance and physiological rPPG cues. It measures cross-domain generalization and compares separate versus joint multi-task learning protocols.

Datasets

• SiW — total ?; splits: train (-1), test (-1)
• 3DMAD — total ?; splits: train (-1), test (-1)
• HKBU-MarsV2 — total ?; splits: train (-1), test (-1)
• MSU-MFSD — total ?; splits: test (-1)
• 3DMask — total ?; splits: test (-1)
• ROSE-Youtu — total ?; splits: test (-1)
• FaceForensics++ — total ?; splits: train (-1), test (-1)
• DFDC — total ?; splits: test (-1)
• CelebDFv2 — total ?; splits: test (-1)

Metrics

• AUC (primary) — range: percent
  • Area Under the Receiver Operating Characteristic Curve. Measures the probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative instance.
• EER (primary) — range: percent
  • Equal Error Rate, the operating point where False Acceptance Rate equals False Rejection Rate.
• TPR@FPR=10% — range: percent
  • True Positive Rate calculated at a fixed False Positive Rate of 10%.
• TPR@FPR=1% — range: percent
  • True Positive Rate calculated at a fixed False Positive Rate of 1%.

Input / output format

Input: Video sequences of faces (typically >5 seconds) containing either bonafide subjects or attack/forgery samples. Inputs may include raw RGB frames, extracted rPPG signals, MSTmaps, or WaveletMaps.

Output: Binary classification score/probability for bonafide vs. spoof/forgery, optionally with a task identifier (spoofing or forgery) for joint training.

Scoring recipe

def compute_metrics(predictions, labels):
    fpr, tpr, _ = roc_curve(labels, predictions)
    auc = auc(fpr, tpr)
    eer = tpr[np.argmin(np.abs(fpr - tpr))]
    tpr_10 = interp(0.10, fpr, tpr)
    tpr_1 = interp(0.01, fpr, tpr)
    return {'AUC': auc, 'EER': eer, 'TPR@FPR=10%': tpr_10, 'TPR@FPR=1%': tpr_1}

Common pitfalls

• Cross-domain evaluation strictly uses entirely unseen datasets for testing without any training data from those domains, making it a zero-shot domain-shift test rather than standard cross-validation.
• Joint training shares parameters across two distinct tasks (physical spoofing vs. digital forgery), which can lead to task interference if modality biases are not explicitly mitigated.
• rPPG extraction is highly sensitive to lighting conditions and head motion, often requiring specific preprocessing that varies across datasets and can skew physiological cue performance.

Evidence (verbatim from paper)

For all experiments, Area Under Curve (AUC) and Equal Error Rate (EER) are utilized for performance evaluation. Besides, the evaluations under intra- and cross-domain scenarios among multiple (merged) datasets are investigated by using the metric of single-side True Positive Rate (TPR)@False Positive Rate (FPR), which is more suitable for realistic spectacles.

Citation

@misc{yu2022benchmarking,
  title={Benchmarking Joint Face Spoofing and Forgery Detection with Visual and Physiological Cues},
  author={Yu et al. (2022)},
  year={2022},
  note={arXiv:2208.05401}
}

• arXiv: 2208.05401