Enhancing Accuracy in Image Recognization for Deepfake Images Using Convolutional Networks Comparing with K-Nearest Neighbors (KNN) Algorithm
C. S., R Surendran, N Madhusundar. · 2025
The primary goal of this research is to improve the accuracy of deepfake picture recognition using Convolutional Neural Networks (CNN) and compare the results to a K-Nearest Neighbors (KNN) approach. This entails researching approaches like transfer learning, feature extraction, and model evaluation to increase the robustness and efficiency of deepfake detection systems. For the training and evaluation phases, a Kaggle dataset with both authentic and deepfaked photos was used. The dataset included samples from several deepfake generating techniques and scenarios. This work involved the implementation of two independent groups, each applying different algorithmic approaches for picture identification. Group 1 used Convolutional Neural Networks (CNNs), while Group 2 used a neural network design with the K-Nearest Neighbors (KNN) algorithm. Each group had 20 samples, for a total sample size of 40. This sample size was chosen to ensure enough statistical power for subsequent studies. Clincalc.com was used for a priori power analysis, and the statistical power (G-power) was set at 80%. The chosen significance level (alpha,$\boldsymbol{\alpha}$) was 0.05, and the type II error rate (beta,$\boldsymbol{\beta}$) was 0.2, A 95% confidence interval was selected to assess the precision of the estimated effectiveness of the two algorithmic approaches. Result: The accuracy of the K-Nearest Neighbors (KNN) algorithm was 72.8%, whereas the accuracy of the Convolutional Neural Networks (CNNs) was 97.15%. A two-tailed significance test confirmed the superiority of the CNN model over the KNN model, with a statistically significant p-value of 0.001 (p 0.05) for accuracy.