Locate
MTCNN extracts and aligns the most prominent face from an uploaded image, or from up to 15 evenly sampled frames of an uploaded video.
01 / Introduction
Detection without explanation isn't good enough anymore.
Deepfake detectors are everywhere, but most of them are black boxes — they return a probability and nothing else. XAI Forensics was built to close that gap: a full-stack system that classifies uploaded images and videos as real or fake, and then shows exactly which region of the face drove that decision.
The project pairs a fine-tuned EfficientNet binary classifier with MTCNN face detection and Grad-CAM explainability on the backend, delivered through a Flask API and a React/TanStack Start forensic dashboard designed to feel like an actual investigative tool rather than a demo toy.
02 / The problem
A label without a reason isn't evidence.
Most deepfake classifiers output a single confidence score and stop there. That's a dead end for journalists, forensic analysts, and platforms that need to justify a decision, not just receive one — a black-box "94% fake" verdict can't be cross-examined, cited, or trusted at face value.
The goal was to build a system that keeps the accuracy of a modern CNN classifier while adding a second, independent layer: visual proof of where the model is looking, so a human reviewer can agree or disagree with the reasoning, not just the label.
03 / The solution
Detect the face, classify it, then explain the classification.
Rather than treating detection as a single opaque step, the pipeline separates face localization, binary classification, and explainability into three distinct stages — so each one can be inspected, tested, and improved independently.
Classify
A EfficientNet backbone, fine-tuned with a custom dropout-linear-ReLU head, scores each detected face and averages the results across all sampled frames for a single verdict.
Explain
Grad-CAM hooks into the final convolutional layer of EfficientNet to generate a heatmap over the highest-confidence face, overlaid and returned alongside the verdict.
SOLUTION / PIPELINE FLOWUPLOAD → MTCNN FACE CROP → EFFICIENTNET SCORE → GRAD-CAM HEATMAP → VERDICT
04 / Model architecture & inference
How a single upload becomes a verdict with visual evidence.
Video-aware, not just image-aware
Images are analyzed directly; videos are decoded and evenly sampled at up to 15 frames using OpenCV, so a short clip and a longer one both get proportional coverage without processing every frame.
MTCNN with margin-padded crops
Each sampled frame is passed through MTCNN with a 20px margin around the detected face, producing consistently sized 224×224 crops that feed directly into the classifier's expected input shape.
Averaging across frames, not just one
Each face crop gets an independent sigmoid probability from EfficientNet; the final verdict is the mean across all detected faces, giving a video-level judgment instead of a single fragile frame-level guess.
05 / Build process
From a trained checkpoint to a live, explainable API.
01Model loadingEfficientNet checkpoint loaded with a custom classification head, with automatic CUDA/CPU fallback.
→
02Face pipelineMTCNN wired in for consistent face alignment on both images and sampled video frames.
→
03ExplainabilityCustom Grad-CAM class registered on EfficientNet's final conv block, with gradient-enabled inference for the top-scoring face.
→
04API + dashboardFlask
/analyze endpoint wired to a React drag-and-drop upload flow with live pipeline-stage animation.BUILD PROCESS / MODEL LOADING → INFERENCE → GRAD-CAM → API RESPONSEPYTORCH → FLASK → REACT
06 / The dashboard
A forensic interface, not just an upload box.
Built with React, TanStack Start, and Framer Motion, the interface walks the reviewer through each pipeline stage in real time — frame extraction, spatial scan, confidence fusion — before surfacing the verdict, spatial and temporal sub-scores, and the Grad-CAM heatmap overlaid directly on the uploaded media.
07 / Limitations and honest framing
What "explainable" does and doesn't mean here.
Heuristic, not a dedicated network
The "temporal anomaly" score shown in the dashboard is derived from the variance of per-frame predictions rather than a purpose-built temporal model like an LSTM or 3D-CNN — a reasonable proxy, but not true motion-consistency analysis.
Illustrative, not independently detected
The listed "findings" (lighting inconsistency, lip-sync drift, frequency artifacts) are selected from a fixed set based on the confidence score rather than each being independently verified by a dedicated detector — useful context, but not standalone evidence.
One face per frame, by design
MTCNN is configured to detect only the most prominent face per frame, so multi-person videos are evaluated on one subject at a time rather than analyzing every face present.
08 / Outcomes and reflections
A working system where the verdict comes with proof, not just a percentage.
XAI Forensics runs end to end: a real upload triggers real MTCNN face detection, real EfficientNet inference, and a real Grad-CAM heatmap — not mocked data. That heatmap is the core deliverable of the project, turning an opaque classifier into something a reviewer can actually inspect and argue with.
The clearest next step is replacing the heuristic temporal score with a genuine sequence model (e.g. a frame-level LSTM or 3D-CNN) and validating the classifier against a public benchmark like FaceForensics++ or Celeb-DF to quantify accuracy beyond the current qualitative testing.
Project details
Built as a NUML final year project, available on GitHub.
PythonPyTorchFlaskMTCNN / facenet-pytorchGrad-CAMReact / TanStack Start
View the project on GitHub ↗
Let's build the next system.
Tell us what you are trying to make clearer, faster, or more capable.
Start a conversation ↗