DEEPFAKE FORENSICS & ANOMALY ANALYSIS WORKBENCHWACV 2025 SPEC

Differential Anomaly Detection (DAD) • Gaussian Mixture Model (GMM) Real-Only Manifold Density

BACKBONE: EFF_B4_SAMGMM COMPONENTS: K=3 DIAG

INGESTION & PARAMETER CONTROL

KEYBIND: ⌘+U

DROP MEDIA FILE OR BROWSE

MP4, MOV, MKV, JPG, PNG (MAX 200MB)

PIPELINE FORENSIC PARAMETERS

DECISION THRESHOLD:0.40

GMM DENSITY MANIFOLD:3-COMP DIAG

GRAD-CAM SPATIAL HEATMAP OVERLAY:

READY FOR PAYLOAD

3-STAGE FORENSIC PIPELINE DIAGRAM

REAL-TIME FLOW

EfficientNet-b4 SAM Backbone

4-Region Convex Hull Pseudo-Deepfake Self-Blending

Real-Only GMM Feature Space

3-Comp Density Fit Exclusively on Real-Real Pairs (A-B)²

Differential Anomaly Score + Grad-CAM

Negative Log-Likelihood -log P(x) + Layer Heatmaps

MATHEMATICAL PREMISE & CITATION

DiffGuard rejects traditional binary classification paradigms. Standard classifiers learn specific generator artifacts and overfit to known deepfake tools.

S_video = (1/M) * ∑ -log P_GMM( (f_A - f_B)² )

Reference: "DiffFake: Exposing Deepfakes using Differential Anomaly Detection" (IEEE/CVF WACV 2025).

Loading telemetry database...