Trident: Detecting Face Forgeries With Domain-Adversarial Triplet Learning

Summary

Trident is a face forgery detection framework that uses triplet learning in a Siamese network to generalize to forgery methods it was not trained on. It adds domain-adversarial training with a forgery discriminator to learn forgery-agnostic embeddings, and blocks gradient flow from the classifier head to the embedding model to reduce overfitting to forgery-specific artifacts. The authors report effectiveness across multiple benchmarks and ablation studies.