Model Card

MAPF-Lite leverages a highly optimized, multi-stream architecture designed for CPU-efficient inference:

• Visual Backbone: Frozen CLIP ViT-B/32. Extracts semantic visual features.
• Audio Backbone: Frozen Whisper Base. Extracts semantic auditory features.
• Forgery Signature Gate: Trainable module utilizing Spatial Rich Model (SRM) filters and 2D Discrete Cosine Transform (DCT) to capture low-level artifact traces (noise inconsistencies, blending boundaries).
• Lite Prompt Generator: Dynamically generates manipulation-method-aware prompts to guide the model's attention.
• Fusion Mechanism: Frame-Level Fusion with Cross-Modal Feature Matching (CMFM) to align audio and visual representations and detect temporal desynchronization.

While highly accurate, MAPF-Lite has known bounds of operational reliability:

• Demographic Variance: Performance may exhibit minor variance across different ethnic groups, skin tones, or ages, largely inherited from the pre-training distributions of the backbone models.
• Environmental Factors: Extreme lighting conditions, significant physical occlusions, or heavy makeup can degrade visual feature extraction.
• Compression: Severe video compression artifacts (e.g., highly compressed social media re-uploads) can destroy the low-level noise patterns relied upon by the Forgery Signature Gate.
• Legal Standing: Output scores are probabilistic. The model is intended as a triage and investigative tool, not as sole, definitive evidence in legal proceedings.