"DEEP FAKE DETECTION”

Mr. SANJAY M 1, P B SANDEEP 2, RACHANA G C 3, SOORYA PRAKASH S 4, SRUJAN S R5

1 Assistant. Professor, Dept. of Information Science & Engineering, Rajeev Institute of Technology, Hassan
2 Information Science & Engineering, Rajeev Institute of Technology,

Hassan 3 Information Science & Engineering, Rajeev Institute of

Technology, Hassan 4 Information Science & Engineering, Rajeev Institute
of Technology, Hassan 5 Information Science & Engineering, Rajeev Institute
of Technology, Hassan

---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Authenticity of Smart Media A method called their temporal correlations. The suggested solution uses CNNs to
Deep Fake identification With Machine Learning uses deep extract significant features from videos and LSTM to classify the
learning approaches to enhance the identification of AI- sequential frame data. This method increases identification
manipulated media. Artificial intelligence (AI) produces accuracy across a range of video formats and situations,
incredibly lifelike synthetic movies known as "deep fakes," automates the deepfake detection process, and lessens the need
which can cause political instability, disinformation, and harm for human verification.
to one's reputation. This project uses preprocessing methods Grade III:AI-powered deepfake detection systems are
like face cropping and frame extraction to analyse video essential for preserving digital integrity and confidence in media
material. While LSTM is used for temporal sequence modelling communications, even after they have first identified
to categorise movies as real or deepfake, ResNeXt CNN is manipulated media. These tools help content platforms,
employed for feature extraction. Real-time detection and journalists, and investigators confirm the authenticity of videos
increased accuracy are the outcomes of the system's automation by detecting fake content. Ensuring the validity of video
of video forensics. It guarantees dependable results and offers evidence and recorded encounters can also be advantageous for
users an easy-to-use online interface by utilising deep learning. the legal, healthcare, and educational sectors. A larger ecosystem
of safe information exchange and digital responsibility is
Key words: LSTM, deepfake detection, facial recognition, supported by the use of deep learning into media forensics.
video forensics, computer vision, deep learning, ResNeXt,
and media authenticity. Grade IV: Deepfake detection using machine learning also
offers wider social and financial benefits. Real-time content
verification and broad accessibility are made possible by
1. INTRODUCTION incorporating detection techniques into mobile applications and
web platforms. This makes digital safety technologies more
Deepfake technology, in which artificial intelligence (AI) accessible to everyone, especially in underserved or distant
algorithms are used to make synthetic yet remarkably realistic regions. Transparency and dependability in digital media are
videos, has emerged as a result of the quick development of improved by the automation and digitisation of detection
deep learning and processing capacity. Because they allow systems, which also expedites legal and regulatory procedures.
media manipulation for nefarious objectives like identity theft, Integrating these solutions into communication platforms as
political propaganda, extortion, and disinformation, these technology advances guarantees a safe online environment for
deepfakes represent a serious threat. In order to differentiate coming generations.
authentic information from fake media, the method presented in
this paper uses sophisticated machine learning algorithms to Media security systems must incorporate machine
analyse and categorise videos. The suggested method improves learning to counter the growing risks of synthetic material. For
the accuracy of deepfake detection by utilising CNN and LSTM
models for frame-level feature extraction and temporal pattern precise, real-time deepfake identification, this study offers an
identification. automated and user-friendly method based on ResNeXt and
LSTM. By reducing human error and improving media
Grade I:In today's media-driven environment, deepfakes verification, the system encourages safe communication and
have emerged as a serious digital danger. Every year, millions well-informed choices. Cross-platform scalability, smartphone
of edited movies are disseminated online, impacting people, integration, and browser plug-ins are possible future
organisations, and governments. Traditional detection
developments that will expand the use and significance of
techniques rely on inefficient and error-prone manual
inspection. Deepfake creation tools and platforms are becoming deepfake detection technologies.
more and more common, hence automated detection techniques
utilising artificial intelligence are crucial. Large video datasets 2. LITERATURE REVIEW
can be analysed using deep learning to find minute irregularities
that indicate manipulation, which helps stop the spread of false A literature review is an essential step in the software
information. development process since it provides valuable insights and
Grade II: Using deep learning models like ResNeXt and improvements for existing methods. This section highlights the
LSTM greatly improves the precision and effectiveness of key papers that have influenced the suggested work on deepfake
deepfake detection. In order to identify face-swapped detection utilising machine learning and deep learning
information, these models examine individual video frames and approaches.
| Page 1
 ultrasound segmentation, and clinical data analysis is needed to
A technique for identifying face warping artefacts in deepfake increase accuracy, efficiency, and early PCOS identification.
videos was presented by Li and Lyu (2018). To find
discrepancies between the created face regions and the Proposed system:
surrounding areas, they employed a convolutional neural
network (CNN). The study highlighted how resolution The suggested method improves the precision and effectiveness
constraints in current deepfake algorithms frequently create of deepfake video detection by utilising machine learning and
artefacts, but it lacked temporal frame analysis, which is deep learning. It combines a Long Short-Term Memory (LSTM)
essential for thorough identification. network to examine temporal inconsistencies and sequential
patterns across video frames with a pre-trained ResNeXt
Since many synthetic videos were unable to replicate realistic Convolutional Neural Network (CNN) for frame-level feature
blink frequencies, Li, Chang, and Lyu (2018) conducted another extraction. To make sure that only pertinent visual information is
significant study that concentrated on identifying deepfakes handled, the system uses a structured pipeline that begins with
using eye-blinking patterns. Using Long-term Recurrent data preprocessing, which includes face detection, frame
Convolutional Networks (LRCN), the model investigated extraction, and noise reduction.
temporal discrepancies in eye behavior. But as deepfake By spotting minute irregularities and artefacts that are not
generation techniques improved, eye-blinking by itself was no evident to the naked eye, this hybrid architecture enables precise
longer enough for accurate detection. categorisation of films as either real or deepfake. To enhance
generalisation across various deepfake creation techniques, the
In order to detect manipulated images and videos in a variety of model is trained on a balanced, varied dataset that incorporates
contexts, including replay attack detection and synthetic face information from several public sources. Users can upload films
recognition, Nguyen et al. (2019) investigated capsule and get real-time forecasts with confidence scores using a web-
networks. Although the method produced encouraging results, based front-end. Future improvements will include real-time
its practicality was restricted by the use of random noise in detection and alert systems that integrate with web browsers and
training. In order to achieve robustness and generalisability, our messaging platforms, guaranteeing greater accessibility and
approach trains on clean, realistic datasets. defence against online disinformation.

In order to detect deepfake videos, Güera and Delp (2018)

suggested using Recurrent Neural Networks (RNNs) in 4. METHODOLOGY
conjunction with CNN-extracted features and temporal
sequence learning. A tiny, homogeneous dataset limited the
model's performance despite its positive outcomes. For real-
time relevance, on the other hand, our system is trained on a
vast and varied dataset that combines many sources.

The FaceForensics++ dataset was first presented by Rossler et

al. (2019) and served as the basis for numerous detection
methods. Their research highlighted the value of carefully
selected, superior datasets for creating reliable classifiers and
showed how well deep learning works to identify edited video.

3. SYSTEM DESIGN
Existing system:

Conventional methods for identifying deepfake movies

mostly depend on forensic video analysis and manual
observation, where professionals spot irregularities like
abnormal lighting and reflections, improper lip-syncing, or
strange facial motions. This approach is subjective, time-
consuming, and prone to human mistake, even if it can be
helpful in certain situations. This is particularly true when
dealing with extremely realistic AI-generated content.
Furthermore, temporal irregularities across frames are
frequently missed by current methods that rely on simple
frame-based comparisons or shallow machine learning models,
which lowers detection accuracy.

Several modern techniques that use semi-automated

ultrasonic analysis but lack advanced image processing and
feature extraction reduce accuracy. The absence of machine
learning-based predictive models in existing diagnostic Figure 1: Architectural Diagram
methods further restricts personalised risk assessment.
Therefore, an AI-driven system that integrates deep learning,
| Page 2
The architecture of a deepfake video detection system is Workshop on Information Forensics and Security
depicted in this flow diagram. A video is first uploaded, and
(WIFS), Hong Kong, China, 1-7.
then preprocessing operations including trimming, facial
detection, and frame extraction are performed. Preprocessed [3] Nguyen, H. H., Yamagishi, J., & Echizen, I. (2019).
data is divided into training and testing sets during the training
Using Capsule Networks to Detect Forged Images
phase, and ResNeXt is used to extract features. LSTM is then
utilised for temporal video classification. A confusion matrix is and Videos. 2019 IEEE International Conference
used to assess the model before it is exported for real-time
on Acoustics, Speech and Signal Processing
forecasting. The trained algorithm uses learnt patterns to
determine whether submitted movies are real or false during (ICASSP), Brighton, UK, 2307-2311.
prediction.
[4] Güera, D., & Delp, E. J. (2018). Deepfake Video
5. CONCLUSIONS Detection Using Recurrent Neural Networks. 2018
15th IEEE International Conference on Advanced
The Deepfake Detection System is a cutting-edge tool
that precisely detects AI-generated fake videos by applying Video and Signal Based Surveillance (AVSS),
deep learning and machine learning algorithms. Using a hybrid Auckland, New Zealand, 1-6.
architecture that incorporates Long Short-Term Memory
(LSTM) networks for temporal video analysis and ResNeXt [5] Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C.,
for frame-level feature extraction, the system successfully Thies, J., & Nießner, M. (2019). FaceForensics++:
identifies movies as real or fraudulent. The trained model's
durability in identifying slight modifications is demonstrated Learning to Detect Manipulated Facial Images.
by its excellent accuracy of up to 97.76% on benchmark IEEE Transactions on Pattern Analysis and
datasets. The solution improves the integrity of digital content,
guarantees prompt intervention, and slows the spread of false Machine Intelligence, arXiv:1901.08971.
information by automating the detection process. [6] Li, Y., Yang, X., Sun, P., Qi, H., & Lyu, S. (2020).
Celeb-DF: A Large-scale Challenging Dataset for
6. FUTURE DIRECTIONS DeepFake Forensics. IEEE Conference on
Computer Vision and Pattern Recognition (CVPR),
Future developments could greatly increase the
Deepfake Detection System's accuracy, scalability, and Seattle, USA, 3207–3216.
accessibility on a variety of platforms. Real-time video
authentication for social media users and content producers
can be made possible by creating small, mobile, and browser-
based plug-ins. Before it spreads, altered content can be
identified and flagged with the aid of integration with
messaging apps and streaming services. For legal and
investigative reasons, tele-forensics capabilities can facilitate
remote collaboration and verification. Furthermore, adding
multimodal analysis—like consistency checks for text and
audio—can increase the system's functionality. Future
developments might potentially include real-time detection of
synthetic voice and full-body deepfakes, guaranteeing
complete defence against manipulation of digital media.
.

REFERENCES

[1] Li, Y., & Lyu, S. (2018). Exposing DeepFake

Videos by Detecting Face Warping Artifacts. 2018
IEEE Conference on Computer Vision and Pattern
Recognition Workshops (CVPRW), Salt Lake City,
USA, 46-52.
[2] Li, Y., Chang, M.-C., & Lyu, S. (2018). In Ictu
Oculi: Exposing AI Created Fake Videos by
Detecting Eye Blinking. 2018 IEEE International

| Page 3
| Page 4