Unmasking Deep Fakes: Leveraging Deep Learning for Video Authenticity Detection
Mahmudul Hasan · 2025
Read Unmasking Deep Fakes: Leveraging Deep Learning for Video Authenticity Detection by Mahmudul Hasan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06528.txt.
Introduction
Mahmudul Hasan and co-authors build a deepfake video detector in a 2025 paper. Fake videos made with AI grow more convincing each year, which threatens trust in online media. Spotting them calls for methods that catch small flaws the eye can miss.
The system first finds faces in each frame with a standard face finder. A second network, trained on images, then judges whether the face looks real or made. Deep learning suits the task because it learns fine patterns on its own. The paper sets out the training and tests for the paired system, aimed at tools that check video truth at scale.
Catalog record
| CREATOR | Mahmudul Hasan |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-06528
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Unmasking Deep Fakes: Leveraging Deep Learning for Video Authenticity Detection Mahmudul Hasan Abstract Deepfake videos, produced through advanced artificial intelligence methods now-a-days, pose a new challenge to the truthfulness of the digital media. As Deepfake becomes more convincing day by day, detecting them requires advanced methods capable of identifying subtle inconsistencies. The primary motivation of this paper is to recognize deepfake videos using deep learning techniques, specifically by using convolutional neural networks. Deep learning excels in pattern recognition, hence, makes it an ideal approach for detecting the intricate manipulations in deepfakes. In this paper, we consider using MTCNN as a face detector and EfficientNet-B5 as encoder model to predict if a video is deepfake or not. We utilize training and evaluation dataset from Kaggle DFDC. The results shows that our deepfake detection model acquired 42.78% log loss, 93.80% AUC and 86.82% F1 score on kaggle’s DFDC dataset. 1 Introduction The swift proliferation of smartphones and social media has resulted in an exponential growth in the creation and sharing of digital images and videos. Alongside this, advancements in artificial intelligence and computer vision have made sophisticated image and video maneuver methods more popular than ever before [ 12 ] . In the past, such modifications required extensive expertise, but with the advent of deep learning-based tools, even individuals with minimal technical knowledge can now alter digital content effortlessly [ 14 ] . One such emerging technology is Deepfake, an artificial intelligence-based method that generates highly realistic but synthetic images, videos, and audio [ 1 ] by manipulating facial expressions or swapping identities seamlessly [ 5 ] . While Deepfake technology has promising applications in entertainment, filmmaking, and gaming, it also raises suggestive moral and security issues [ 10 ] . The capability to create convincingly fabricated media has fueled the spread of misinformation, political disinformation, and identity fraud. Malicious actors exploit Deepfake technology to generate deceptive content, leading to reputational damage, blackmail, and cyber harassment [ 15 ] . A notable example is the case of journalist Rana Ayyub, whose manipulated video went viral in 2018, causing severe personal and professional harm [ 18 ] . The increasing misuse of Deepfake technology poses a serious threat to digital media authenticity and national security, with reports indicating a 67% rise in security breaches related to manipulated media. Deepfake technology primarily relies on deep learning models, especially Generative Adversarial Networks (GANs), to synthesize hyper-realistic content that is often imperceptible from genuine media [ 2 ] . Although these techniques enhance visual effects in various industries, their potential for misuse necessitates the development of robust detection frameworks. Recent studies suggest that nearly 50% of the billions of media files uploaded daily on online platforms undergo some form of manipulation [ 4 ] . To counteract this growing issue, researchers have explored various detection methodologies, including AI-driven forensic analysis and blockchain-based authentication mechanisms such as Smart Contracts and Hyperledger Fabric [ 16 ] . However, despite these efforts, Deepfake detection remains a challenging and evolving field that requires continuous improvement. This paper intent to enhance the accuracy of deepfake detection by integrating two powerful neural network models: Multi-task Cascaded Convolutional Networks (MTCNN) for facial identification and EfficientNet for image classification. By leveraging a large and high-quality dataset, we hypothesize that our approach will effectively distinguish between authentic and manipulated media, as indicated by a low binary cross-entropy value. Through further optimization and testing, we anticipate that our framework will significantly improve the reliability of Deepfake detection, offering a robust solution to mitigate the risks associated with this technology. The key contributions of this paper are as follows: • We proposed a MTCNN-EfficientNetB5 fusion model to detect deepfake videos where MTCNN was utilized for face detection and EfficientNetB5 for feature extraction and classification. • We have tested our model on the DFDC dataset where the proposed model outperformed many state-of-the art models without extensive data processing and ensembling models. The remainder of this paper is structured as follows: Section 2 reviews prior research on Deepfake detection methods. Section 3 describes the dataset used in this study. Section 4 outlines the proposed methodology, including data preprocessing steps, model architectures and training. Section 5 presents the performance evaluation metrics and results analysis. Section 6 concludes the paper. 2 Related Work Despite the availability of various deepfake detection methods, identifying deepfakes remains a significant challenge. This is primarily due to the sophisticated nature of deep neural networks, which make it difficult to trace manipulation artifacts [ 13 ] . Additionally, celebrities are frequent targets of deepfake manipulations, including face-swapping and lip synchronization, which current detection techniques often fail to accurately detect. To overcome these obstacles, researchers have looked into strategies like dual-stream learning methods to improve detection capabilities and protect individuals from malicious deepfake applications [ 8 ] . The increasing accessibility of cloud computing has further exacerbated the deepfake problem, as it enables individuals with no prior knowledge of computer vision or deep learning to generate highly realistic fake videos. Consequently, numerous detection models have been proposed, including transfer learning, autoencoders, and hybrid models combining CNNs and RNNs [ 17 ] . One notable detection framework, DFN, utilizes MobileNet as its base architecture, incorporating separable convolution and max-pooling layers. DFN has demonstrated superior performance compared to Xception and EfficientNet, indicating its potential for enhancing deepfake detection [ 6 ] . In addition to AI-driven methods, biometric-based forensic techniques have arrived as a auspicious solution for deepfake detection. These approaches leverage two key methodologies: Static Biometric Analysis, which employs traditional facial recognition techniques to analyze still images in videos, and Temporal Behavioral Analysis, which examines facial movements and expressions over time to detect inconsistencies. Such methods enhance detection accuracy by considering both spatial and temporal aspects of deepfake content [ 2 ] . Deng, Suo, and Li’s study, "Deep Fake Video Detection Based on EfficientNet-V2 Network," proposes an innovative technique that integrates longitudinal data from video sequences with the EfficientNet-V2 architecture to improve detection performance [ 11 ] . While their approach shows promise in enhancing feature extraction and detection efficiency, it lacks comprehensive validation in real-world scenarios. Tan and Le’s research, "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," introduces a systematic method for scaling CNNs to achieve state-of-the-art performance while optimizing computational resources [ 19 ] . Their EfficientNet architecture employs a compound scaling algorithm that uniformly adjusts model depth, width, and resolution. This method enables EfficientNet to outperform traditional architectures such as ResNet and Inception in image classification tasks while requiring fewer computational resources. However, a limitation of this approach is its reliance on predefined scaling values, which may not be universally optimal. Additionally, its primary application to image classification limits its effectiveness for tasks involving spatial analysis, such as object detection. Maryam Taeb and Hongmei Chi’s study, "Comparison of Deepfake Detection Techniques through Deep Learning," evaluates various deep learning architectures for deepfake detection, with an emphasis on robustness against fraud, reliability, and computational efficiency [ 18 ] . Their work highlights the advantages and limitations of different models, including CNNs and RNN-based hybrid approaches, across multiple datasets. While the study provides valuable insights into best practices for deepfake detection, it lacks extensive video-based analysis and