Deepfake detection problem from R



Working with video datasets, notably with respect to detection of AI-based pretend objects, could be very difficult because of correct body choice and face detection. To method this problem from R, one could make use of capabilities supplied by OpenCV, magick, and keras.

Our method consists of the next consequent steps:

  • learn all of the movies
  • seize and extract pictures from the movies
  • detect faces from the extracted pictures
  • crop the faces
  • construct a picture classification mannequin with Keras

Let’s rapidly introduce the non-deep-learning libraries we’re utilizing. OpenCV is a pc imaginative and prescient library that features:

Alternatively, magick is the open-source image-processing library that may assist to learn and extract helpful options from video datasets:

  • Learn video recordsdata
  • Extract pictures per second from the video
  • Crop the faces from the pictures

Earlier than we go into an in depth rationalization, readers ought to know that there is no such thing as a must copy-paste code chunks. As a result of on the finish of the publish one can discover a hyperlink to Google Colab with GPU acceleration. This kernel permits everybody to run and reproduce the identical outcomes.

Information exploration

The dataset that we’re going to analyze is supplied by AWS, Fb, Microsoft, the Partnership on AI’s Media Integrity Steering Committee, and varied teachers.

It accommodates each actual and AI-generated pretend movies. The full dimension is over 470 GB. Nevertheless, the pattern 4 GB dataset is individually out there.

The movies within the folders are within the format of mp4 and have varied lengths. Our process is to find out the variety of pictures to seize per second of a video. We often took 1-3 fps for each video.

Word: Set fps to NULL if you wish to extract all frames.

video = magick::image_read_video("aagfhgtpmv.mp4",fps = 2)
vid_1 = video[[1]]
vid_1 = magick::image_read(vid_1) %>% image_resize('1000x1000')

We noticed simply the primary body. What about the remainder of them?

Trying on the gif one can observe that some fakes are very simple to distinguish, however a small fraction appears fairly reasonable. That is one other problem throughout information preparation.

Face detection

At first, face places have to be decided by way of bounding packing containers, utilizing OpenCV. Then, magick is used to mechanically extract them from all pictures.

# get face location and calculate bounding field
unconf <- ocv_read('frame_1.jpg')
faces <- ocv_face(unconf)
facemask <- ocv_facemask(unconf)
df = attr(facemask, 'faces')
rectX = (df$x - df$radius) 
rectY = (df$y - df$radius)
x = (df$x + df$radius) 
y = (df$y + df$radius)

# draw with crimson dashed line the field
imh  = image_draw(image_read('frame_1.jpg'))
rect(rectX, rectY, x, y, border = "crimson", 
     lty = "dashed", lwd = 2)

If face places are discovered, then it is rather simple to extract all of them.

edited = image_crop(imh, "49x49+66+34")
edited = image_crop(imh, paste(x-rectX+1,'x',x-rectX+1,'+',rectX, '+',rectY,sep = ''))

Deep studying mannequin

After dataset preparation, it’s time to construct a deep studying mannequin with Keras. We will rapidly place all the pictures into folders and, utilizing picture mills, feed faces to a pre-trained Keras mannequin.

train_dir = 'fakes_reals'
width = 150L
top = 150L
epochs = 10

train_datagen = image_data_generator(
  rescale = 1/255,
  rotation_range = 40,
  width_shift_range = 0.2,
  height_shift_range = 0.2,
  shear_range = 0.2,
  zoom_range = 0.2,
  horizontal_flip = TRUE,
  fill_mode = "nearest",

train_generator <- flow_images_from_directory(
  target_size = c(width,top), 
  batch_size = 10,
  class_mode = "binary"

# Construct the mannequin ---------------------------------------------------------

conv_base <- application_vgg16(
  weights = "imagenet",
  include_top = FALSE,
  input_shape = c(width, top, 3)

mannequin <- keras_model_sequential() %>% 
  conv_base %>% 
  layer_flatten() %>% 
  layer_dense(items = 256, activation = "relu") %>% 
  layer_dense(items = 1, activation = "sigmoid")

mannequin %>% compile(
  loss = "binary_crossentropy",
  optimizer = optimizer_rmsprop(lr = 2e-5),
  metrics = c("accuracy")

historical past <- mannequin %>% fit_generator(
  steps_per_epoch = ceiling(train_generator$samples/train_generator$batch_size),
  epochs = 10

Reproduce in a Pocket book


This publish exhibits how one can do video classification from R. The steps had been:

  • Learn movies and extract pictures from the dataset
  • Apply OpenCV to detect faces
  • Extract faces by way of bounding packing containers
  • Construct a deep studying mannequin

Nevertheless, readers ought to know that the implementation of the next steps might drastically enhance mannequin efficiency:

  • extract the entire frames from the video recordsdata
  • load completely different pre-trained weights, or use completely different pre-trained fashions
  • use one other expertise to detect faces – e.g., “MTCNN face detector”

Be happy to strive these choices on the Deepfake detection problem and share your leads to the feedback part!

Thanks for studying!


For those who see errors or wish to recommend modifications, please create a difficulty on the supply repository.


Textual content and figures are licensed beneath Inventive Commons Attribution CC BY 4.0. Supply code is out there at, until in any other case famous. The figures which were reused from different sources do not fall beneath this license and will be acknowledged by a word of their caption: “Determine from …”.


For attribution, please cite this work as

Abdullayev (2020, Aug. 18). RStudio AI Weblog: Deepfake detection problem from R. Retrieved from

BibTeX quotation

  writer = {Abdullayev, Turgut},
  title = {RStudio AI Weblog: Deepfake detection problem from R},
  url = {},
  yr = {2020}


Leave a Reply

Your email address will not be published. Required fields are marked *