Deepfake technology, driven by advancements in deep learning and generative models, enables highly realistic

manipulation of facial appearances in videos, often through face-swapping techniques. While such methods

have potential in entertainment and creative applications, they also pose serious threats to privacy, trust, and

information integrity. This paper presents the development of a machine learning (ML)-based system for

detecting face-swap deepfake videos. The proposed approach employs video preprocessing, frame extraction,

and facial region isolation, followed by feature extraction using a deep convolutional neural network

(ResNeXt). Temporal consistency is analyzed with a Long Short-Term Memory (LSTM) network to capture

particular, replace the identity of one person with another using advanced generative models like Generative

Adversarial Networks (GANs) and autoencoders. These manipulations are often visually convincing, making

them difficult to detect with the naked eye. While deepfake technology offers creative possibilities in film

production, virtual reality, and education, it also introduces risks such as political misinformation,

cyberbullying, and identity theft. The rapid evolution of deepfake generation techniques has led to increasingly

realistic outputs, challenging traditional detection methods. Manual verification is both time-consuming and

unreliable, especially in large-scale media platforms where content is uploaded in massive volumes daily.[2]

The detection of deepfakes has thus emerged as a critical research area, requiring robust, automated, and

scalable solutions capable of adapting to evolving forgery techniques. This work proposes an ML-based

detection framework capable of identifying subtle visual and temporal inconsistencies in manipulated content,

researchers examined pixel-level inconsistencies, unnatural facial boundaries, and blending artifacts that often

occurred due to imperfect alignment and color matching between the source and target faces. Afchar et al.

convolution-based generative networks fail to reproduce the natural spectral distribution of real images, making

al. (2018), leveraged biological cues like eye-blinking patterns, pulse signals, and head movements, which are

often missing, irregular, or temporally inconsistent in deepfakes. With the advent of high-quality manipulations

speech—have shown promise in catching advanced deepfakes.[5]Despite these advancements, challenges

remain, particularly in handling deepfakes that are adversarially trained to bypass detection, and maintaining

that can provide interpretable results for both forensic analysis and public trust.[6]

The origins of face-swapping can be traced back to the late 1990s and early 2000s, when photo editing

software such as Adobe Photoshop introduced tools capable of replacing facial regions in static images. These

early techniques relied on manual editing and image morphing, often producing results that were far from

realistic. In the late 2000s, advancements in computer vision and 3D modeling enabled more automated face

hardware. Around the same time, the term deepfake—a combination of “deep learning” and “fake”—was

coined on an online forum, marking the beginning of a surge in publicly shared AI-generated videos, many

and detection. Social media platforms and AI communities began experimenting with commercial applications,

but malicious uses such as political misinformation campaigns, identity theft, and non-consensual explicit

content raised serious ethical and legal concerns. In response, tech companies and research institutions initiated

Today, face-swap deepfakes continue to evolve, with improvements in model architecture, training data

diversity, and post-processing making detection increasingly difficult, thus driving ongoing research in

enhances data quality and ensures uniformity across samples.[9]

Together, these features form a robust input representation for classification.

usability and accessibility of the deepfake detection framework in real-world scenarios.[10]

The results obtained from the proposed deepfake detection model demonstrate its effectiveness in accurately

distinguishing between real and manipulated videos. This section presents a detaile evaluation of model

performance, training behavior, and comparative analysis with existing approaches.[11]

Fig I. Image of real video Detected

Fig III. Result of training model

accuracy improves steadily from around 49% to 96%, while the validation accuracy increases from 67% to

nearly 99%, indicating effective learning. Simultaneously, both training and validation loss values decrease

Fig. IV Evaluation of the training model

The model achieved 99% accuracy on the test set with a very low loss (0.0701), indicating excellent

performance. This suggests it generalizes well to unseen data.

Table I. Performance Evaluation of the Proposed Deepfake Detection Model

Comparison with Other

XceptionNet)

Better accuracy & scalability

Outperformed existing detection

Key Strengths

Detects lighting inconsistencies, unnatural

expressions, and abrupt transitions

Effective in identifying deep fake

manipulations

The proposed ML-based solution effectively identifies face-swap deepfakes by combining spatial and temporal

Additionally, exploring hybrid approaches that combine audio-visual cues, physiological signal analysis, and

explainable AI methods will help in building a more transparent, trustworthy, and comprehensive deepfake

detection framework.[13]

1. Goodfellow, I., et al. (2014). Generative adversarial nets. Advances in Neural Information Processing

2. Afchar, D., Nozick, V., Yamagishi, J., & Echizen, I. (2018). MesoNet: a compact facial video forgery

3. Li, Y., Chang, M.C., & Lyu, S. (2018). In Ictu Oculi: Exposing AI created fake videos by detecting eye