ANALYZING VISUAL FEATURES FOR REAL VS. FAKE FACE DETECTION: PREPARING FEATURE-AWARE HYBRID CNN MODELS

Abstract

Recent advances in generative adversarial networks (GANs) have led to the availability of high-quality synthetic facial images (deepfakes) at large scale. These include but are not limited to, such advances pose challenges in digital security, misinformation detection and biometric identification. The above methods are used in this work to combat the growing threat of the developing synthetic media by understanding the discriminative features between actual and fake face images. The primary goal is to discover and learn the most discriminative handcrafted feature, which we can use to build more efficient hybrid CNN-based classifiers in subsequent studies. We show a variety of methods of feature extraction and visualization by leveraging public dataset “140k Real and Fake Faces”. Color histograms, Local Binary Patterns (LBP), Sobel, Canny edge detectors are some of the traditional image descriptors for the representation of color, texture, edge and frequency data. We have applied Discrete Cosine Transform (DCT). These manually created features have been converted into 2-D spatial models by PCA, tSNE and UMAP to describe their ability to separate classes. This work further pinpoints significant cues for distinguishing real faces from computer-generated or printed ones, and lays the groundwork for a hybrid CNN model that integrates learned and hand-crafted features.