/Deepfake_with_tensor_decomposition

Primary LanguageJupyter Notebook

Deepfake with tensor decomposition

Summer 2020, Internship at DASH Laboratory, SKKU

Searching for Deepfake Detection Algorithm using Tensor Decomposition

Dive into Dataset

Experiments

Configurations

  • Dataset Composition
REAL FAKE(NT) RATIO
Total Dataset 10000 10000
Training Dataset 8100 8100
Validation Dataset 900 900
Test Dataset 1000 1000
  • Training Options
Options Values
Batch size 32
Epochs 100
Base Network XceptionNet
Initialization imagenet weights
Augmentations rescale=1./255
rotation_range=20
width_shift_range=0.1
height_shift_range=0.1
shear_range=0.1
zoom_range=0.1
horizontal_flip=True
fill_mode='nearest'
Optimizer Adam
loss function Binary Cross Entropy
metrics accuracy
recall
precision
f1
Callbacks checkpointing
cyclical LR-exp_range(1e-3 ~ 7e-3)

Baseline

https://github.com/dongminkim0220/Deepfake_with_tensor_decomposition/blob/master/tensor_dcmp_baseline.ipynb

Precision Recall F1
REAL 0.78 0.94 0.85
FAKE 0.92 0.74 0.82
AUROC 0.943088
THRESH 0.1833098829
TOTAL ACCURACY 0.839

baseline accuracy baseline loss

TK with rank = [30, 30, 3]

https://github.com/dongminkim0220/Deepfake_with_tensor_decomposition/blob/master/tensor_dcmp_TK.ipynb

Precision Recall F1
REAL 0.77 0.94 0.85
FAKE 0.92 0.71 0.81
AUROC 0.940521
THRESH 0.0516793727880219
TOTAL ACCURACY 0.8285

tk accuracy tk loss

TK_diff

https://github.com/dongminkim0220/Deepfake_with_tensor_decomposition/blob/master/tensor_dcmp_TK_diff.ipynb

Precision Recall F1
REAL 0.76 0.6 0.67
FAKE 0.67 0.81 0.74
AUROC 0.78239
THRESH 0.628564715385419
TOTAL ACCURACY 0.708

tk diff accuracy tk diff loss