Towards Intrinsic Common Discriminative Features Learning for Face Forgery Detection using Adversarial Learning

• Realistic face forgeries create security and privacy risks, motivating stronger techniques for detecting fake faces.Source: §1 · p. 1
• Xception features can cluster by forgery method and facial identity, showing that vanilla classification learns information beyond real/fake labels.Source: §1 · p. 1, Figure 1 · p. 1
• Identity-related semantic features and method-specific texture patterns can limit generalization across identities and manipulation algorithms.Source: §1 · pp. 1–2
• The method uses adversarial learning with forgery-method and identity discriminators to eliminate these factors and learn intrinsic common discriminative features.When facial identity labels are unavailable, its identity discriminator uses similarity information from an off-the-shelf face recognition model.Source: §1 · p. 2

• Early methods detected artifacts such as reduced blinking frequency or visual inconsistencies in eyes, noses, and teeth.Source: §2.1 · p. 2
• Methods targeting common manipulation steps, including affine warping and blending, can suffer performance drops on low-quality images because compression destroys artifact cues.Source: §2.1 · p. 2
• Data-driven deep networks achieved high intra-dataset accuracy but did not account for extra dataset information unrelated to real/fake classification.Source: §2.1 · p. 2
• Unlike frequency-clue and attention-based approaches, this method uses adversarial learning to eliminate forgery-method and facial-identity effects.Source: §2.1 · p. 2, §2.2 · p. 2

Overview

• Adversarial learning makes the feature representation indistinguishable across forgery methods and facial identities while retaining features for real/fake classification.A feature generator is paired with forgery-method and identity discriminators, while a binary classifier predicts whether each face is real or fake.Source: §3.1 · p. 2, Figure 2 · p. 3, §3.4 · p. 4

3.2. Forgery Method Adversarial Learning

• The forgery-method discriminator classifies manipulation types, whereas the feature generator is optimized oppositely so its representation captures common properties across forgeries.The discriminator minimizes adversarial loss and the generator maximizes it, reducing method-specific information in the extracted features.Source: §3.2 · pp. 2–3

3.3.1. Hard Identity Label Supervision

• The identity discriminator suppresses facial-identity information using adversarial training, with hard identity labels applicable when such labels are available.For face swapping, the fake face receives the target identity label; for facial reenactment, it retains the source identity label.Source: §3.3 · p. 3, §3.3.1 · p. 3

3.3.2. Identity Similarity Supervision

• When identity labels are unavailable, ArcFace embeddings provide cosine-similarity supervision that is thresholded into binary labels with τ = 0.07.The identity discriminator compares elementwise squared differences between generated feature embeddings and outputs a scaled similarity prediction.Source: §3.3.2 · p. 3

Experimental Setup

• The experiments use FaceForensics++ for intra-dataset evaluation and train on it for cross-dataset tests on Celeb-DF and DeepFakeDetection.Table 1 reports video-level results across quality settings, while Table 2 reports frame-level AUC for cross-dataset evaluation.Source: §4.1.1 · p. 4, Table 1 · p. 4, Table 2 · p. 4

Cross-Dataset Results

• On Celeb-DF, training on FaceForensics++ yields superior performance to other methods except Two-branch, despite using no temporal information.The method is only slightly inferior to Two-branch, which uses temporal information.Source: §4.1.2 · p. 4, Table 2 · p. 4

Component Effectiveness

• Adversarial learning disperses features across forgery methods and identities while making real and fake feature distributions more compact.The t-SNE analysis attributes this pattern to eliminating forgery-method and facial-identity effects.Source: §4.2.1 · p. 5, Figure 3 · p. 5
• Both adversarial components improve intra-dataset detection and cross-dataset generalization over the baseline, while their combination produces further gains.The results identify forgery-method and identity adversarial learning as complementary components.Source: §4.2.1 · p. 5, Table 3 · p. 5

Generalization to Unseen Forgery Methods

• The proposed method consistently outperforms the baseline when tested on unseen FaceForensics++ forgery methods after training on the remaining four.The evaluation covers DeepFakes, Face2Face, FaceSwap, NeuralTextures, and FaceShifter in the low-quality setting.Source: §4.2.2 · p. 5, Table 4 · p. 5

Identity Supervision

• Identity similarity supervision outperforms hard identity-label supervision, whereas pseudo-label supervision performs worse because clustering is inaccurate.The authors attribute hard-label errors to incorrectly assigning swapped fake faces to target identities.Source: §4.2.3 · p. 6, Figure 4 · p. 6

• Adversarial learning targets the effects of forgery methods and facial identities to learn intrinsic common discriminative features.Source: §5 · p. 6
• Two adversarial discriminators address forgery methods and facial identities, producing a feature space independent of both factors.Source: §5 · p. 6
• A similarity-based identity discriminator uses an off-the-shelf face recognition model when facial-identity ground-truth labels are unavailable.Source: §5 · p. 6
• Experiments evaluate the proposed method in both intra-dataset and cross-dataset settings.Source: §5 · p. 6

• Xception maps input images into 2048-D embedding feature vectors in the latent representation space.Source: §A.1 · p. 7
• The binary classifier uses a fully connected layer with two output nodes representing real versus fake.Source: §A.1 · p. 7
• The forgery-method discriminator uses three fully connected layers with 512, 512, and N units for N manipulation methods.Source: §A.1 · p. 7

• Detected faces are cropped around the face center with 1.3× enlargement and resized to 299 × 299.Source: §A.2 · p. 7
• Face detection and landmark localization use the open-source dlib algorithm.Source: §A.2 · p. 7
• The feature generator is initialized with ImageNet-pretrained parameters.Source: §A.2 · p. 7
• Training uses batch size 64, a maximum of 20,000 iterations, and Adam with initial learning rate 0.0001.Source: §A.2 · p. 7

• Lower clustering accuracy is used to indicate weaker correlation between learned features and the clustered attribute.Source: §B.1 · p. 7
• Table 5 reports forgery-method clustering accuracy for four FaceForensics++ manipulation methods.Source: Table 5 · p. 7
• Table 6 reports identity clustering accuracy for 28 facial identities from DeepFakeDetection original videos.Source: Table 6 · p. 7
• Compared with the baseline, clustering accuracy drops 7.72% with K-Means and 8.54% with GaussianMixture.The reported reductions concern the clustering analyses used for forgery-method and identity associations.Source: §B.1 · p. 7

Adversarial loss weights λ1, λ2

• Best detection accuracy occurs with λ1=0.8 and λ2=5 on the FaceForensics++ low-quality task.Performance rises and then falls as the adversarial-loss weights vary.Source: §B.2.1 · p. 7
• The identity-similarity threshold is selected by estimating training-data cosine-distance distributions from 100 randomly sampled batches.The cumulative distribution of within-batch facial cosine distances is used to limit the threshold range.Source: §B.2.2 · p. 8, Figure 6 · p. 8
• The best FaceForensics++ low-quality video-level performance is achieved when τ=0.07.The threshold is evaluated across a series of τ values.Source: §B.2.2 · p. 8, Figure 7 · p. 8