Deepfake Detection using Multi-path CNN and Convolutional Attention Mechanism
Rineesh Babu P., Prof. (Dr.) Madhu S. Nair · 2022 IEEE 2nd Mysore Sub Section International Conference (MysuruCon) · 2022
Image and video forgery using cutting-edge deep learning techniques has become one of the major issues in the social networking era. Media manipulation in which one person’s face is swapped out for another’s or has additional features added is referred to as deepfakes. Despite the fact that it has many beneficial purposes, fraudsters generally utilise it to create celebrity porn, revenge porn, and fake news, among other things. One of the biggest risks that deepfake presents is that people’s belief in the reality of many things may decline. The motivation behind deepfake detection is the need to prove that the real thing is real and the fake thing is fake. In this paper a multi-CNN approach for detecting deepfakes is being proposed. Here, a multipath convolutional neural network (CNN) with three modules is used, each of which is stacked with a convolutional block attention mechanism. The first two modules in the dual-path paradigm are a Resnet module and a Densenet module. The Resnet component enables for feature reuse while Densenet allows for the investigation of new features. The parallel InceptionResnet module contains a one-dimensional feature reduction module with residual connections. When the performance of the proposed model is compared with that of four deep learning based approaches, it is found that the proposed method gave the best outcomes, with an accuracy and F1-score of 0.940 and 0.939, respectively.