shairkhanfaizan/deepverify-models

DeepVerify V1 Models

DeepVerify is an AI-generated/manipulated media detection project. This repository contains the V1 production model bundle, utilizing a 60/40 spatial-frequency probability ensemble.

Models Included

1. Spatial Model (EXP-001)

   • Architecture: EfficientNet-B0
   • Representation: Spatial RGB
   • Weight: 60% of ensemble

2. Frequency Model (EXP-002)

   • Architecture: EfficientNet-B0
   • Representation: Frequency-domain log-magnitude FFT
   • Weight: 40% of ensemble

Input & Output

• Expected Input: 224x224 RGB image
• Output Interpretation: Both models predict P(REAL). The final ensemble score is 0.60 * Spatial_P(REAL) + 0.40 * Frequency_P(REAL). A score >= 0.5 is considered LIKELY_AUTHENTIC, while < 0.5 is considered LIKELY_AI_GENERATED.

Limitations

• Dataset Limitations: Trained exclusively on the 140k Real and Fake Faces v1.0 benchmark. Does not guarantee out-of-distribution generalization to all generative models or non-face semantics.
• Model Limitations: This model is not a universal deepfake detector. It is highly specific to the domain it was trained on.
• Confidence: The probability score is a statistical confidence metric, not absolute ground truth.
• Explainability: Spatial Grad-CAM visual evidence indicates regions of model activation/saliency. It is not causal proof of manipulation.