xdfdet — Explainable Deepfake Detection

This repository contains the code and experiments for regional explainability analysis of CNN-based deepfake detectors. Rather than proposing a new detection architecture, this work investigates what facial regions drive model decisions under different preprocessing configurations, using Grad-CAM activation mapping across eight landmark-defined facial areas.

Explainable Deepfake Detection Using Frame Level CNN Models: A Comparative Study of Augmentation and Cutout Techniques
Mert Kaya — MSc Computer Engineering, TED University, 2025
Supervisor: Asst. Prof. Dr. Venera Adanova
DOI: 10.5281/zenodo.18998566

Reference implementation. The thesis linked above is the authoritative and complete description of the methodology, configurations, and results — including the full true/false positive and negative regional activation analyses and the per-region standard-deviation analysis. The notebooks here provide the core training and explainability pipeline; for the exact hyperparameters and the complete set of figures, refer to the thesis.

Motivation

Deepfake detection models regularly report AUC scores above 0.90, yet they remain largely unusable in contexts where decisions must be justified. A binary "real/fake" output with a confidence score is not sufficient when:

• Legal proceedings require evidence that can be examined and challenged. A judge or forensic analyst cannot cross-examine a black-box prediction. Detection outputs need to be traceable to specific visual cues in order to qualify as admissible evidence.
• Content moderation at scale flags millions of items. When a piece of content is removed, the creator is entitled to know why. "The model said so" is not an acceptable explanation under transparency requirements like the EU AI Act and DSA.
• Model auditing and failure analysis is impossible without understanding where the model focuses. If a detector achieves 90% AUC but relies on JPEG compression artifacts rather than facial manipulation cues, it will fail silently on re-compressed or high-quality deepfakes.

Existing benchmarks (DeepfakeBench, DeepFake-o-Meter) evaluate detection accuracy across multiple models and datasets, but none of them analyze the spatial attention patterns that underlie those predictions. This work addresses that gap.

Approach

The detection pipeline is based on Seferbekov's DFDC competition solution, adapted for FaceForensics++:

1. Frame extraction — 12 frames per video at equal intervals, face-cropped to 224×224 using MTCNN
2. SSIM-guided cutout — structural similarity between real and fake frames identifies the most similar (hardest to detect) facial region; a polygon cutout is applied to fake frames with black, white, or random fill
3. Star cutout on real frames — small star-shaped masks prevent overfitting to pristine facial texture
4. Augmentation — Albumentations pipeline (noise, blur, geometric, color) at two intensity levels
5. Classification — EfficientNet-B4 backbone in a TimeDistributed wrapper, frame-level sigmoid predictions averaged for video-level classification
6. Explainability — Grad-CAM heatmaps computed per frame, mapped onto eight dlib 68-point landmark regions (left/right eyes, eyebrows, nose, inner/outer mouth, jaw), aggregated across TP/TN/FP/FN cases

Nine configurations are compared, varying augmentation (on/off, low/high probability) and cutout (on/off, black/white/random fill). All other hyperparameters are held constant.

Findings

Augmentation without cutout degrades performance. The augmentation-only model scored lower than the no-augmentation baseline (AUC 0.8610 vs 0.8678). Combining augmentation with SSIM-guided cutout recovered and exceeded baseline performance (0.8971). This suggests that augmentation introduces noise that is only beneficial when paired with targeted regularization.

Fill strategy has a measurable effect. Black-filled cutout with augmentation achieved the highest AUC (0.8971), followed by random (0.8837) and white (0.8734). This contradicts Seferbekov's original finding that random fill performs best — likely due to the difference in dataset scale (1000 pairs here vs 100K+ in DFDC).

Different metrics produce different rankings. The augmentation-only model ranks last in AUC but second in F1-score. The random-filled model ranks second in AUC but first in LogLoss. Evaluating deepfake detectors on a single metric can be misleading.

The nose region dominates model attention regardless of correctness. Grad-CAM regional analysis shows that the nose has the highest mean activation across all four prediction categories (TP, TN, FP, FN) and across all nine configurations. The model over-relies on nose-region features for both correct and incorrect decisions.

Augmentation produces more distributed attention patterns. Models trained with augmentation show activation standard deviation of 11–27% across facial regions, compared to 25–37% without. More distributed attention correlates with lower false positive rates, while concentrated attention on a single region leads to overfitting.

Results

Model	AUC	F1	Brier	LogLoss
blackfilledwithaug	0.8971 ± 0.0064	0.8429 ± 0.0101	0.1242 ± 0.0009	0.4710 ± 0.0063
randomfilledwithaug	0.8837 ± 0.0072	0.7950 ± 0.0098	0.1450 ± 0.0011	0.4656 ± 0.0067
whitefilledwithaug	0.8734 ± 0.0061	0.7951 ± 0.0095	0.1455 ± 0.0012	0.4761 ± 0.0071
noaugrandomfill	0.8711 ± 0.0051	0.7769 ± 0.0069	0.1524 ± 0.0012	0.5241 ± 0.0078
onlymoreaug	0.8718 ± 0.0043	0.7932 ± 0.0077	0.1431 ± 0.0008	0.5288 ± 0.0081
baseline	0.8678 ± 0.0044	0.7780 ± 0.0065	0.1524 ± 0.0013	0.4827 ± 0.0069
noaugblackfill	0.8666 ± 0.0043	0.7912 ± 0.0077	0.1537 ± 0.0009	0.4989 ± 0.0074
noaugwhitefill	0.8639 ± 0.0047	0.7704 ± 0.0074	0.1462 ± 0.0011	0.5244 ± 0.0076
onlyaug	0.8610 ± 0.0051	0.7955 ± 0.0084	0.1572 ± 0.0010	0.5719 ± 0.0084

All models trained 3 times on FaceForensics++ (1000 video pairs, 70/15/15 split). Values are mean ± std.

Limitations

• Dataset is limited to 1000 video pairs from FF++. Results may differ at larger scale.
• No cross-dataset evaluation (CelebDF, DFDC).
• Frame-level spatial analysis only — no temporal modeling (LSTM, attention over frames).
• Single backbone architecture (EfficientNet-B4).

Reproduction

Two paths are supported.

Option A — reproduce from scratch (train).

1. Place the FaceForensics++ videos under DATA_ROOT (see 01_data_preparation). The same subset as the thesis is used: 1000 real/fake pairs across FaceSwap, Face2Face, FaceShifter, and Deepfakes, split 70/15/15.
2. Run 01_data_preparation and 02_preprocessing_pipeline to build the pairs and the preprocessing functions.
3. Run 03_training with USE_PRETRAINED = False to train the selected ACTIVE_CONFIG. Weights are written to weights/<config>.h5.
4. Run 04_evaluation and 05_explainability.

Option B — use the pretrained weights.

1. Download the trained weights from:
2. In 03_training, set USE_PRETRAINED = True and either set WEIGHTS_URL to the direct download link or place the .h5 file at weights/<config>.h5.
3. Run 01_data_preparation and 02_preprocessing_pipeline (needed to build the test set and preprocessing functions).
4. Run 03_training (it loads the weights instead of training), then 04_evaluation and 05_explainability.

FaceForensics++ is distributed under its own license; obtain dataset access from the dataset authors.

Repository Structure

notebooks/
  01_data_preparation.ipynb      # FF++ pairing, splitting
  02_preprocessing_pipeline.ipynb # SSIM, landmarks, cutout, augmentation
  03_training.ipynb               # 9 configs, model, training
  04_evaluation.ipynb             # AUC, F1, LogLoss, Brier, curves
  05_explainability.ipynb         # Grad-CAM, regional analysis, case comparison

app/
  app.py                          # Streamlit demo (in development)

Setup

Python 3.9+, TensorFlow 2.10+, GPU recommended.

pip install tensorflow albumentations scikit-image scikit-learn opencv-python dlib seaborn tqdm

dlib requires shape_predictor_68_face_landmarks.dat — download from dlib.net.

Notebooks are designed for Google Colab with GPU runtime. Run 01–04 sequentially to reproduce training. Notebook 05 requires a trained model.

Citation

@mastersthesis{kaya2025xdfdet,
  title     = {Explainable Deepfake Detection Using Frame Level CNN Models:
               A Comparative Study of Augmentation and Cutout Techniques},
  author    = {Kaya, Mert},
  school    = {TED University},
  year      = {2025},
  doi       = {10.5281/zenodo.18998566}
}

License

MIT