Deepfake authenticity evaluation studio
A controlled evaluation studio for comparing deepfake detectors across prepared image and video datasets, compression levels, and manipulation types.
Project definition
Problem statement
Deepfake detectors can perform well on the dataset used for training but fail on a different generator, compression level, resolution, or editing pipeline. A confidence score alone cannot prove that media is authentic or manipulated.
The engineering problem is to compare detectors under controlled conditions, measure cross-condition robustness and calibration, and show the evidence and limits behind each result.
Project objectives
- Prepare versioned real and manipulated media splits with documented source and consent status.
- Run selected detectors through a common frame, clip, and media-level interface.
- Compare performance across manipulation type, compression, resolution, and unseen-source splits.
- Measure confidence calibration and define an uncertain range for human review.
- Display frame-level evidence and common failure cases without claiming certainty.
System design
System modules
Media registry
Stores source, label, consent or licence, manipulation type, quality, split, and file hashes.
Preprocessing pipeline
Extracts metadata, samples frames, detects faces where required, and creates fixed media inputs.
Detector runner
Executes selected model adapters and records frame, clip, and media-level scores and runtime.
Robustness laboratory
Applies prepared compression, resizing, and transcoding conditions and reruns the same evaluation.
Evidence dashboard
Shows scores, calibration, frame timelines, explanations, errors, and review status.
Methodology
System workflow
- 01Prepare dataset
Permitted media is hashed, labelled, documented, and assigned to fixed source-separated splits.
- 02Configure experiment
The student selects detector versions, preprocessing, quality conditions, and metrics.
- 03Run detection
Each detector processes the same prepared media and records predictions and performance.
- 04Test robustness
Controlled compression and resizing variants are evaluated against the original results.
- 05Review failures
The dashboard groups false results and uncertain cases by source and media condition.
Demonstration scenario
Two detectors process a prepared set containing real and manipulated videos from separated sources. The dashboard compares their media-level results, then repeats the test after compression. One detector loses accuracy, and the student inspects the frame timeline and places an uncertain case into human review.
Engineering
Technical architecture
- Web application
- Nuxt and Vue for dataset control, experiment setup, metric comparison, evidence timelines, and review.
- Evaluation API
- FastAPI for media jobs, detector adapters, robustness variants, predictions, metrics, and exports.
- Data layer
- PostgreSQL for media metadata, dataset versions, models, experiments, predictions, metrics, and reviews.
- Media pipeline
- Python, OpenCV, and FFmpeg for controlled decoding, frame selection, transformations, and metadata capture.
- Model evaluation
- PyTorch detector adapters with source-separated tests, threshold analysis, calibration, and saved model versions.
Testing
Evaluation
Evaluation measures
- Area under the ROC and precision-recall curves
- False-positive and false-negative rates at selected thresholds
- Performance on unseen manipulation sources
- Accuracy loss under compression, resizing, and transcoding
- Calibration error and selective accuracy for reviewed cases
- Inference time and failure rate by media duration and quality
System boundaries
- Only licensed, consented, public research, or safely generated media is included.
- The result is an experimental detector score and not proof that media is authentic or fake.
- The system does not identify the person shown or infer intent, authorship, or criminal activity.
- Real decisions require provenance checks, source investigation, and qualified human review.
Included
- 01Media dataset and experiment-management interface
- 02Detector comparison, robustness, and calibration pipelines
- 03Frame evidence and error-analysis dashboard
- 04Prepared licensed media, experiments, and evaluation results
- 05Complete source code in a private GitHub repository
- 06Complete project documentation with synopsis, abstract, methodology, model diagrams, experiment results, screenshots, and conclusion
- 07Setup and usage guide
Project record
No buyer information is collected on this page.
- Permanent project ID
- GP-CY-0W0GOCJ
- Catalogued
- 21 Aug 2026
- Completed
- Pending
- Verified
- Pending
- Demonstration
- Added when ready
Handover
After purchase
- 01Payment is confirmed
The project is marked unavailable and cannot be purchased again.
- 02Repository access is granted
The buyer's submitted GitHub account receives access to the private repository.
- 03The purchase record is delivered
The certification sheet is prepared from the reviewed buyer details and sent privately by email.