Naveenck10/MediaDNA-V22.4F

MediaDNA V22.4F

This repository hosts the frozen research checkpoints for the MediaDNA V22.4F multimodal media authenticity prototype.

Model Checkpoints

• Primary checkpoint: V22_4F_MULTIMODAL_StageB_Ep2.pth
• Primary architecture: VideoCAVMAEFT / AVFF-family multimodal detector

The V22.4F model is the PRIMARY detector.

The following are AUXILIARY DIAGNOSTIC specialists:

• V22_3B_VISUAL_CHECKPOINT.pth
• V22_3C_AUDIO_CHECKPOINT.pth

Scientific Output Definition

The primary V22.4F model produces a RAW SIGMOID DECISION SCORE.

This score is NOT a calibrated probability.

Frozen operating policy:

• score < 0.20 → AUTHENTIC
• 0.20 <= score <= 0.45 → UNCERTAIN
• score > 0.45 → SYNTHETIC

Note: Do not call the score a probability, calibrated probability, or confidence probability.

Frozen Results

The frozen locked-test results (evaluated on 2,219 samples) are exactly:

• ROC-AUC: 0.7500
• PR-AUC: 0.9899
• Balanced Accuracy: 0.7686
• MCC: 0.1679
• Precision: 0.9956
• Recall: 0.6621
• Specificity: 0.8750
• F1: 0.7953

Confident confusion matrix:

• TN = 42
• FP = 6
• FN = 697
• TP = 1366

Important observed weakness: FakeVideo + RealAudio recall = 23.16%. (Do not describe this checkpoint as universally reliable.)

Limitations

• This is a research/engineering prototype.
• Evaluation is tied strictly to the tested dataset/protocol; external generalization is not established.
• The raw sigmoid score is not a calibrated probability.
• Model-sensitive evidence is not ground-truth manipulation localization.
• Exact generator attribution is unsupported.
• FakeVideo + RealAudio performance is substantially weaker.
• Outputs should not be treated as definitive legal/authenticity proof.