Detection of Facial Manipulated Images using Convolutional Neural Networks
Abstract
Recent advances in Generative Adversarial Network (GAN)-based image synthesis and autoencoder-driven facial manipulation have raised serious concerns about the trustworthiness of visual media. This paper presents a compact yet effective deepfake detection framework built on a Convolutional Neural Network (CNN) trained to distinguish authentic facial images from GAN-manipulated counterparts sourced from the FaceForensics++ benchmark. The detection scope is bounded to GAN- and autoencoder-based facial manipulations; diffusion-model-generated images fall outside the present evaluation corpus and represent a direction for future work. The system employs a three-stage preprocessing pipeline: Haar-cascade facial detection to isolate forgery-susceptible regions; per-image min-max intensity normalization to remove sensor-dependent exposure offsets; and stochastic online data augmentation to mitigate overfitting to compression artefacts. The CNN comprises ten layers-three convolutional, three max-pooling, one flatten, one fully connected with dropout, and one Softmax output-yielding 6,423,042 trainable parameters and trained with categorical cross-entropy loss. Experiments on a balanced $\mathbf{1 4, 0 0 0}$-image corpus yield 91.3% test accuracy and an F1-score of 0.912. Five-fold cross-validation across demographically diverse subsets returns a mean accuracy of 91.1% ± 0.5%. Comparison against seven published methods confirms statistically significant gains, and the complete pipeline achieves end-to-end inference within 200 ms on a commodity GPU.