Enhancing Accuracy in Image Recognition for Deepfaked Images using Convolutional Networks Compared with Random Forest Algorithm
C. S., R Surendran, S Raveena · 2025
The primary goal of this study is to improve the accuracy of deepfake image recognition using convolutional neural networks (CNN) and compare the results to Random Forest approaches. Deepfake detection systems can increase their dependability and success by experimenting with strategies such as feature extraction. transfer learning, and model evaluation. Materials and Methods: This is a Kaggle dataset. A combination of actual and deep-faked images is utilized for training and assessment. The dataset contains samples created under various conditions and approaches for deepfake generation. This work implemented two distinct groups, each employing a different algorithmic technique for image identification. Group 1 employed CNNs (convolutional neural networks). Group 2 used the random forest algorithm with a hierarchical tree-like topology. There were forty samples in all, with twenty in each group. A sample size was determined to provide adequate statistical power for future studies. An initial power calculation was conducted using Clincalc.com; the statistical power (G-power) was set at 80%. Behavioural science conventions let one find the type II error rate (beta) and significance level (alpha). The values chosen were 0.2 and 0.05, respectively. The accuracy of the two algorithmic techniques was examined using a 95% confidence interval. Result: The Random Forest had an accuracy of 71.28%, whereas Convolutional Neural Networks (CNNs) beat it with a 97.15% accuracy. A two-tailed significance test result for accuracy showed The CNN model has a statistically significant p-value of 0.001 (p < 0.05). This suggests that the observed variations in accuracy between the two models are statistically significant. Discussions: Convolutional networks outperformed other models, which was explained by their ability to extract features from images