Detecting Deepfake Videos: An Analysis of Three Techniques

View on arXiv ← Back to list

Authors: Armaan Pishori, Brittany Rollins, Nicolas van Houten, Nisha Chatwani, Omar Uraimov

Published: 2020-07-15 20:36:23+00:00

AI Summary

This paper analyzes three techniques for deepfake video detection: convolutional LSTM, eye blink detection, and grayscale histograms. The authors participated in the Deepfake Detection Challenge and found the grayscale histogram technique to be the most relevant, showing promise for high accuracy with further development.

Abstract

Recent advances in deepfake generating algorithms that produce manipulated media have had dangerous implications in privacy, security and mass communication. Efforts to combat this issue have risen in the form of competitions and funding for research to detect deepfakes. This paper presents three techniques and algorithms: convolutional LSTM, eye blink detection and grayscale histograms-pursued while participating in the Deepfake Detection Challenge. We assessed the current knowledge about deepfake videos, a more severe version of manipulated media, and previous methods used, and found relevance in the grayscale histogram technique over others. We discussed the implications of each method developed and provided further steps to improve the given findings.

Key findings

The grayscale histogram method achieved the highest accuracy (85.71%) and lowest validation loss among the three tested methods. While all models achieved accuracies in the 80-90% range, the results suggest that the grayscale histogram approach has significant potential for improvement with more data and refined techniques. Computational limitations restricted the amount of data used for training and may have influenced the overall accuracy.

Approach

The authors implemented three different approaches: a convolutional LSTM network, eye blink detection using OpenCV and a KNN classifier, and a grayscale histogram approach with an LSTM network. These methods were compared to a baseline CNN+RNN model without preprocessing to evaluate their effectiveness.

Datasets

500 GB of video data from the Kaggle Deepfake Detection Challenge, a subset of approximately 50 GB was used for training.

Model(s)

Convolutional LSTM, KNN, LSTM network, CNN+RNN

Author countries

USA

← Previous