Fast and Generalized DeepFake Detector Through Feature Space Transformation

TMLR Paper3671 Authors

Abstract: The current state-of-the-art DeepFake or manipulated image detection algorithms are not generalized against an unseen database, manipulation types, and image degradation due to compression. Existing literature shows different input transformations can boost the detection performance of deepfake detection algorithms. However, these algorithms only transform the spatial pixel values with the hope that the transformation will help in learning a linearly separable decision boundary. The transformation of a 2D volume containing millions of pixel values is computationally complex and on top of that, the amalgamation with the original image further increases the computational complexity. The proposed algorithm utilizes the concept of transformation; however, the transformation of a feature space that is 1-D and compact representation of an image. The transformed representation is then used to calculate the discriminative feature maps used for the binary classification as real or altered images. Extensive experimentation on multiple databases under several unconstrained settings establishes the effectiveness of the proposed algorithm and its desirability in the current era. Under each set, the proposed algorithm achieves state-of-the-art detection performance on Face Forensics++ and Celeb-DF databases. The proposed algorithm is almost `parameter-free' and achieves its two-fold aim of giving a robust detection algorithm and an energy-saving medium.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Charles_Xu1

Submission Number: 3671