DEEPFAKE DEFENDER – REAL-TIME VIDEO CALL INTEGRITY CHECKER USING DEEP LEARNING AND EXPLAINABLE AI

Abstract— The Deepfake technology has developed
at an incredible speed in the recent years and has become a
threat to our digital security. There are many instances where
we communicate with people in real time and thus have to
worry about their identities being verified. Despite the numerous
techniques and methods that have been proposed to resolve this
deepfake detection problem, none of the current solutions have
proven to be robust enough to work accurately in a real time
scenario, due to inconsistencies in their performance.
In this study, a system for the real-time detection of deepfake
is proposed focusing on image analysis method and relying on
the analysis of audio signals as well. For the analysis of visual
signals, an EfficientNet is used after processing the data and
providing the appropriate training. Audio signals are converted
into Mel- spectrograms and are also analyzed using a CNN to
highlight the characteristics of the voices.
Real time Deepfake Video Detection using CNN with Mel
Spectrogram and Image Pre processing Index Terms—Deepfake
Detection, Efficient Net, CNN, Mel Spectrogram The proposed
system performs video detection of deepfakes in real time. It
captures frames from the webcam of the device during runtime
and classifies whether the input is real or fake. The image model achieves a classification accuracy in between 76–85% and the model performs stably during real time classification. Although it is not a perfect model, it shows that deep learning with both image and audio inputs can be used to strengthen trust in digital communication.