DeepfakeDet-ViT — XAI & Digital Forensics Suite

Drag & Drop Image, Paste, or Browse

Supports JPG, PNG, WEBP (Auto-cropped to 384×384)

Analysis Status

Awaiting Input Image

ViT Attribution Overlays Click any map to inspect in fullscreen modal

Attention Rollout

Chefer Gradient-Rollout

GradCAM

Integrated Gradients

DeepLIFT

Occlusion Sensitivity

1. How to Read XAI Heatmaps

Explainable AI (XAI) methods attribute the final prediction score back to the original image pixels. Warm colors (red/yellow) indicate high attribution, meaning those pixels strongly influenced the model's prediction. Cool colors (blue/purple) indicate low attribution or minimal influence. Each method approaches this problem differently, effectively answering a different question about the model's internal behavior.

2. XAI Methods

Attention Rollout

Averages attention weights across all heads and propagates them through all 12 layers. It highlights where the model focused its spatial attention globally, regardless of the final class prediction.

Chefer Gradient-Rollout

Multiplies attention weights by their backward gradients. This makes the map class-discriminative, showing exactly which regions pushed the probability toward the final verdict (REAL or FAKE).

GradCAM

Uses the gradients of the target concept flowing into the final attention layer. It provides the cleanest, most localized answer to "where did the model decide?", often highlighting the face center.

Integrated Gradients

Computes the integral of gradients along a straight path from a black baseline image to the input. It meticulously captures pixel-level contributions, highlighting fine details that influenced the score.

DeepLIFT

Decomposes the differences in non-linear activations relative to a baseline reference. It is excellent at finding isolated "hotspots" or blending artifacts along face boundaries.

Occlusion Sensitivity

The most faithful, model-agnostic method. It slides a black patch over the image and measures the drop in prediction score. Bright regions directly mean: "if you cover this, the model changes its mind."

3. Classic Forensics Tools

Error Level Analysis (ELA)

Reveals differing compression levels in an image. Spliced or manipulated areas often stand out as noticeably brighter or darker than the surrounding native pixels when re-compressed.

Spatial Edge Gradient

Calculates high-frequency transitions to expose unnatural edge profiles. Face swaps and AI generations frequently leave subtle halos, double edges, or unnatural sharpness at blending seams.

Bit Plane Extractor

Strips away significant visual data to analyze the least significant bits (LSB). AI synthesis often leaves structured, blocky, or geometrically regular noise patterns in the lowest bit planes.

MinMax & Wavelet Estimators

Statistical and frequency domain techniques to find regions where pixel noise variance or intensity drastically deviates from the local neighborhood, typical of pasted imagery.

4. Model Architecture

Model: buildborderless/CommunityForensics-DeepfakeDet-ViT
Architecture: Vision Transformer (ViT-Small). Inputs are resized to 384×384. It contains 12 transformer blocks (layers) and 6 attention heads per layer. The final layer passes through a binary sigmoid classifier. The model was trained on 2.7M samples across 4,803 distinct generative models and manipulation techniques.

DeepfakeDet-ViT XAI & Forensics Suite

ViT Attribution Overlays Click any map to inspect in fullscreen modal

Contribute Ground Truth to Research Dataset

Interactive Method Parameter Explorer

Attention Head Disaggregation (6 Heads per Layer)

Classic Digital Image Forensics Suite

Error Level Analysis (ELA)

1. How to Read XAI Heatmaps

2. XAI Methods

Attention Rollout

Chefer Gradient-Rollout

GradCAM

Integrated Gradients

DeepLIFT

Occlusion Sensitivity

3. Classic Forensics Tools

Error Level Analysis (ELA)

Spatial Edge Gradient

Bit Plane Extractor

MinMax & Wavelet Estimators

4. Model Architecture