NEURAL ARCHITECTURE SPECIFICATION

MAPF-Lite
Forensic Engine.

Method-Aware Prompting and Fusion — a dual-stream, CPU-quantized neural architecture designed for sub-second deepfake detection and transparent attribution.

PIPELINE DECONSTRUCTION

How inference
executes in real time.

Select any stage below to inspect mathematical formulations, CPU tensor shapes, and forensic heuristics.

Stage 01 Specification — FFmpeg Decoded

1. Media Chunking & Audio Demux

V \in \mathbb{R}^{T \times 3 \times 224 \times 224}, \quad A \in \mathbb{R}^{T \times 16000}

Functional Role

Videos are decoded on CPU into 16 evenly spaced representative frames alongside 16kHz mono audio channels.

Forensic Rationale

By uniformly sampling across temporal intervals rather than extracting every frame, CPU memory footprint is capped at under 900MB during active worker runs.

TEMPORAL COHERENCE

Audio-Visual Temporal Radar.

Compare multi-signal coherence between authentic speech and synthetic lip-syncing in real time.

5-Dimensional Audio-Visual Telemetry

Forensic Metrics Data Table

Axis	Score	Description
LipSync Coherence (S_AV)	85%	Cross-modal temporal alignment between speaker phonemes and lip-motion visemes.
Spatial SRM Residuals	88%	Spatial Rich Model 5x5 steganographic filtering detecting seam blending artifacts.
Spectral 2D-DCT Decay	89%	Discrete Cosine Transform decomposition exposing synthetic periodic grid anomalies.
Facial Boundary Consistency	95%	Edge gradient inspection detecting FaceSwap boundary interpolation inconsistencies.
Acoustic Synthesis Markers	84%	Whisper Base auditory prompt conditioning uncovering neural voice-clone timbre artifacts.

Cross-Modal Coherence (S_AV)

0.85 (High)PASS: Coherent

Phonemes extracted from Whisper align smoothly with visual lip visemes computed across 16 frames.

Spatial Boundary Seam Anomaly

0.12 (Clean)NATURAL GRADIENT

Ready to verify suspicious media footage?

Launch Scan Bay View Model Card

Loading DeepShield AI...

MAPF-Lite
Forensic Engine.

How inference
executes in real time.

Media Chunking & Audio Demux

Forgery Signature Gate (SRM + 2D-DCT)

Dual Frozen Pre-trained Backbones

Method-Aware Prompt Tuning

Cross-Modal Fusion & A/V Radar

1. Media Chunking & Audio Demux

Functional Role

Forensic Rationale

Audio-Visual Temporal Radar.

Forensic Metrics Data Table

Ready to verify suspicious media footage?

MAPF-Lite
Forensic Engine.

How inference
executes in real time.

Media Chunking & Audio Demux

Forgery Signature Gate (SRM + 2D-DCT)

Dual Frozen Pre-trained Backbones

Method-Aware Prompt Tuning

Cross-Modal Fusion & A/V Radar

1. Media Chunking & Audio Demux

Functional Role

Forensic Rationale

Audio-Visual Temporal Radar.

Forensic Metrics Data Table

Ready to verify suspicious media footage?

MAPF-Lite Forensic Engine.

How inference executes in real time.

Media Chunking & Audio Demux

Forgery Signature Gate (SRM + 2D-DCT)

Dual Frozen Pre-trained Backbones

Method-Aware Prompt Tuning

Cross-Modal Fusion & A/V Radar

1. Media Chunking & Audio Demux

Functional Role

Forensic Rationale

Audio-Visual Temporal Radar.

Forensic Metrics Data Table

Ready to verify suspicious media footage?

MAPF-Lite Forensic Engine.

How inference executes in real time.

Media Chunking & Audio Demux

Forgery Signature Gate (SRM + 2D-DCT)

Dual Frozen Pre-trained Backbones

Method-Aware Prompt Tuning

Cross-Modal Fusion & A/V Radar

1. Media Chunking & Audio Demux

Functional Role

Forensic Rationale

Audio-Visual Temporal Radar.

Forensic Metrics Data Table

Ready to verify suspicious media footage?

MAPF-Lite
Forensic Engine.

How inference
executes in real time.

MAPF-Lite
Forensic Engine.

How inference
executes in real time.