A Research Paper On Deepfake Face Detection Using Machine Learning
Fuente:
Zenodo
Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Recurso digital |
| Pubblicazione: |
Zenodo
2026
|
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866901193143353344 |
|---|---|
| author | Ashish Nanotkar Sharvari Nachane Lalan Samrit Dewang Pimpalkar Ayush Mallewar |
| author_facet | Ashish Nanotkar Sharvari Nachane Lalan Samrit Dewang Pimpalkar Ayush Mallewar |
| contents | <p>The fast development of Generative AI has resulted in a boom in extremely realistic fake media often called deepfakes and such media represents a serious risk to digital identity to the accuracy of information and to information security. Conventional forensic methods based on metadata examination or visual inspection are becoming less and less able to counteract the advanced pixel level editing and the cutting edge voice synthesis techniques. To fill this gap, in this paper a new multi-modal deepfake detection scheme capable of identifying both visual pipelines and audio signal pipelines is proposed. The architecture employs deep learning and transfer learning in particular XceptionNet for video frames analysis and MobileNetV2 for audio verification in a solid Python Flask based web application. The system employs a two-pipeline methodology: the visual part identifies frames and detects faces using Haar Cascades and pass these frames to the XceptionNet classifier and the acoustic part transforms audio tracks into Mel-spectrograms with Librosa and classifies them with the MobileNetV2 network. This bi-modal processing guarantees the correct detection even if only one of the modalities has been manipulated in the media instance Balanced datasets of authentic and synthetic media are used for experimental evaluation of our framework. We observe a high efficacy of the proposed system with the detection accuracy of around 98% achieved during training and good validation metrics obtained during testing of the system across the two modalities. The system is delivered with a user friendly graphical interface for accessible application in real-world scenarios such as social media moderation, identity verification in digital KYC, and digital legal forensics. By automating the detection and going deeper into the observed metadata, by analyzing the biological and spectral evidence as well, this multi modal framework offers a scalable, energy efficient, high performance solution to a rapidly growing epidemic of the AI generated disinformation.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19816878 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | A Research Paper On Deepfake Face Detection Using Machine Learning Ashish Nanotkar Sharvari Nachane Lalan Samrit Dewang Pimpalkar Ayush Mallewar <p>The fast development of Generative AI has resulted in a boom in extremely realistic fake media often called deepfakes and such media represents a serious risk to digital identity to the accuracy of information and to information security. Conventional forensic methods based on metadata examination or visual inspection are becoming less and less able to counteract the advanced pixel level editing and the cutting edge voice synthesis techniques. To fill this gap, in this paper a new multi-modal deepfake detection scheme capable of identifying both visual pipelines and audio signal pipelines is proposed. The architecture employs deep learning and transfer learning in particular XceptionNet for video frames analysis and MobileNetV2 for audio verification in a solid Python Flask based web application. The system employs a two-pipeline methodology: the visual part identifies frames and detects faces using Haar Cascades and pass these frames to the XceptionNet classifier and the acoustic part transforms audio tracks into Mel-spectrograms with Librosa and classifies them with the MobileNetV2 network. This bi-modal processing guarantees the correct detection even if only one of the modalities has been manipulated in the media instance Balanced datasets of authentic and synthetic media are used for experimental evaluation of our framework. We observe a high efficacy of the proposed system with the detection accuracy of around 98% achieved during training and good validation metrics obtained during testing of the system across the two modalities. The system is delivered with a user friendly graphical interface for accessible application in real-world scenarios such as social media moderation, identity verification in digital KYC, and digital legal forensics. By automating the detection and going deeper into the observed metadata, by analyzing the biological and spectral evidence as well, this multi modal framework offers a scalable, energy efficient, high performance solution to a rapidly growing epidemic of the AI generated disinformation.</p> |
| title | A Research Paper On Deepfake Face Detection Using Machine Learning |
| url | https://doi.org/10.5281/zenodo.19816878 |