Release 8 xdfdet checkpoints with model card

.gitattributes

@@ -33,3 +33,13 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/card.png filter=lfs diff=lfs merge=lfs -text
+assets/gradcam.png filter=lfs diff=lfs merge=lfs -text
+aug-cutout-black.keras filter=lfs diff=lfs merge=lfs -text
+aug-cutout-random.keras filter=lfs diff=lfs merge=lfs -text
+aug-cutout-white.keras filter=lfs diff=lfs merge=lfs -text
+aug-standard.keras filter=lfs diff=lfs merge=lfs -text
+baseline.keras filter=lfs diff=lfs merge=lfs -text
+cutout-black.keras filter=lfs diff=lfs merge=lfs -text
+cutout-random.keras filter=lfs diff=lfs merge=lfs -text
+cutout-white.keras filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,95 @@
+---
+license: cc-by-nc-4.0
+library_name: keras
+pipeline_tag: video-classification
+thumbnail: https://huggingface.co/mertkayacs/xdfdet/resolve/main/assets/card.png
+tags:
+  - deepfake-detection
+  - explainability
+  - grad-cam
+  - efficientnet
+  - faceforensics
+  - tensorflow
+---
+
+# xdfdet: deepfake detectors with Grad-CAM region analysis
+
+Eight EfficientNet-B4 video classifiers from the paper **"Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention"** (UBMK 2026) and the MSc thesis behind it. Each one was trained on FaceForensics++ under a different augmentation and cutout setting, so you can compare what the preprocessing changes, down to which face regions the model looks at.
+
+[Code](https://github.com/mertkayacs/xdfdet) · [Project site](https://xdfdet.mertkayacs.com) · [Thesis](https://doi.org/10.5281/zenodo.18998566)
+
+<img src="assets/gradcam.png" alt="Grad-CAM of the aug-cutout-black model on a public-domain NASA portrait: activation concentrates on the left eye and the nose" width="320">
+
+*Grad-CAM of `aug-cutout-black` on a public-domain NASA portrait. Real-probability 0.9996; strongest activation on the left eye (81%) and the nose (63%).*
+
+## Use
+
+```bash
+pip install git+https://github.com/mertkayacs/xdfdet
+xdfdet predict video.mp4 --gradcam cam.png          # default model: aug-cutout-black
+xdfdet predict video.mp4 --model baseline
+```
+
+```python
+import xdfdet
+model = xdfdet.load_model("aug-cutout-black")
+```
+
+Input is a batch of 12 face crops of 224×224 RGB, normalized with ImageNet mean and std: shape `(batch, 12, 224, 224, 3)`. Output is the probability that each frame is **real**, shape `(batch, 12, 1)`. Average over frames for the video score; below 0.5 means fake. The `xdfdet` package handles face cropping (MTCNN), frame sampling and normalization. Use it rather than feeding raw frames, because the preprocessing has to match training exactly.
+
+## Models
+
+| File | Augmentation | Cutout fill | AUC | F1 | Brier | LogLoss |
+|---|---|---|---|---|---|---|
+| `aug-cutout-black.keras` | standard | black | **0.8981** | **0.8431** | **0.1247** | 0.4710 |
+| `aug-cutout-random.keras` | standard | random | 0.8820 | 0.7950 | 0.1451 | **0.4656** |
+| `aug-cutout-white.keras` | standard | white | 0.8734 | 0.7883 | 0.1451 | 0.4761 |
+| `cutout-white.keras` | flip only | white | 0.8700 | 0.7703 | 0.1463 | 0.5244 |
+| `baseline.keras` | none | none | 0.8684 | 0.7781 | 0.1523 | 0.4827 |
+| `cutout-black.keras` | flip only | black | 0.8669 | 0.7911 | 0.1537 | 0.4989 |
+| `cutout-random.keras` | flip only | random | 0.8642 | 0.7774 | 0.1526 | 0.5241 |
+| `aug-standard.keras` | standard | none | 0.8616 | 0.8025 | 0.1576 | 0.5719 |
+
+Scores are each checkpoint's own run on its 150-pair FaceForensics++ test split, taken from the original training notebook. The paper reports mean ± std over three runs per configuration; those numbers are in the [code README](https://github.com/mertkayacs/xdfdet#results). The ninth configuration, `aug-intense`, is described in the paper, but its checkpoint was lost.
+
+## Training
+
+- **Data:** 1,000 real/fake pairs from FaceForensics++, one fake per real video, with FaceSwap, Face2Face, FaceShifter and Deepfakes in rotation. Split 70/15/15 into train, validation and test.
+- **Preprocessing:** MTCNN face detection with eye alignment, 32 frames per video, 12 used per sample.
+- **Model:** EfficientNet-B4 (ImageNet init) applied to each frame, then global average pooling, dropout 0.55 (0.25 for `aug-cutout-random`), and a sigmoid unit with L2 1e-3.
+- **Loss:** binary cross-entropy on the frame-averaged prediction.
+- **Optimization:** Adam with cosine decay (1e-3 over 1,000 steps) and gradient clipping at 1.0. Batch of 4 pairs, early stopping on validation loss with patience 5, at most 20 epochs.
+- **Environment:** TensorFlow 2.19 / Keras 3.10 on a Colab T4, in mixed precision. The released files are float32 copies with bit-identical weights. Scores can differ from the float16 runs, most near the 0.5 boundary; `conversion.json` lists both for two test images. For the original precision on a GPU, load with `xdfdet.load_model(name, mixed_precision=True)`. `scripts/convert_checkpoints.py` in the code repo shows how each Colab checkpoint was matched to its configuration.
+
+## Limits
+
+- Trained and tested only on FaceForensics++. Expect lower accuracy on other datasets, on newer generators such as diffusion-based face swaps, and on heavily compressed video.
+- Each configuration used its own random split, so a checkpoint may have seen, during training, videos that are in another configuration's test set. Do not re-score these checkpoints on a shared FaceForensics++ split and compare them.
+- Performance across age, sex and skin tone was not measured.
+- These are research models. A score from them is not evidence that a video is real or fake, and it should not be used on its own for decisions about people.
+
+## License
+
+CC BY-NC 4.0. The models were trained on FaceForensics++, whose terms allow non-commercial research and educational use only. The code is MIT.
+
+## Citation
+
+```bibtex
+@inproceedings{kaya2026augmentation,
+  title     = {Augmentation and Cutout in Deepfake Detection: A Comparative Study of
+               Accuracy, Calibration, and Attention},
+  author    = {Kaya, Mert and Adanova, Venera},
+  booktitle = {11th International Conference on Computer Science and Engineering (UBMK 2026)},
+  year      = {2026},
+  note      = {To appear}
+}
+
+@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}
+}
+```

aug-cutout-black.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfb4ad8a40edfc22a75e8910cc28f113b432b80c5ba50647641ad622eb125490
+size 72364819

aug-cutout-random.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3b5e1a3f230f247fb0d53e8f1f840e7024166395f07f95cc5fc9795888f04e1
+size 72364819

aug-cutout-white.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9d10cc5eb99d378da6d29ca3cfdf0564dc83fdc751ba413836c61b4ce4f57e5
+size 72364819

aug-standard.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa5903362048d62e4292cd2f23ff0a201f779a869b45751f3ce7b30e6aaf2251
+size 72364798

baseline.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a79f7560dda198c180929076214b919f85251d3915b21d333ae2e71a6ad6e9fb
+size 72364786

conversion.json

@@ -0,0 +1,98 @@
+{
+  "baseline": {
+    "source": "baseline_no_aug_no_cutout.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      0.468505859375,
+      0.493896484375
+    ],
+    "video_score_release": [
+      0.6664999723434448,
+      0.10909999907016754
+    ]
+  },
+  "aug-standard": {
+    "source": "justaug.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      1.0,
+      1.0
+    ],
+    "video_score_release": [
+      1.0,
+      1.0
+    ]
+  },
+  "cutout-random": {
+    "source": "noaug.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      0.0,
+      0.0
+    ],
+    "video_score_release": [
+      0.0,
+      0.0
+    ]
+  },
+  "cutout-black": {
+    "source": "noaugcb.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      0.0,
+      0.0011997222900390625
+    ],
+    "video_score_release": [
+      0.0,
+      0.007699999958276749
+    ]
+  },
+  "cutout-white": {
+    "source": "noaugcw.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      1.0,
+      1.0
+    ],
+    "video_score_release": [
+      1.0,
+      1.0
+    ]
+  },
+  "aug-cutout-random": {
+    "source": "rc12.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      0.52197265625,
+      0.4453125
+    ],
+    "video_score_release": [
+      0.40709999203681946,
+      0.45080000162124634
+    ]
+  },
+  "aug-cutout-black": {
+    "source": "bzeroc12.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      0.99365234375,
+      0.9990234375
+    ],
+    "video_score_release": [
+      0.9980000257492065,
+      0.9994999766349792
+    ]
+  },
+  "aug-cutout-white": {
+    "source": "whitec12.h5",
+    "weights_identical": true,
+    "video_score_mixed_float16": [
+      1.0,
+      0.9990234375
+    ],
+    "video_score_release": [
+      0.9998999834060669,
+      0.9979000091552734
+    ]
+  }
+}

cutout-black.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d2e5cd08e681c8f9324a6f480d98549b66cffbaf710378581cb982b93d2663b
+size 72364812

cutout-random.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acda23932604cc38fabb217b035617d90c4f7e79185495ad1397e37af06d5524
+size 72364812

cutout-white.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f26b6d96d772b53e2bc8e1abdfb19c275eb93a592cf0c52bb43381424399ba1f
+size 72364812