srirammayan/deepfake-detection

Purpose and Entry Points

This repository implements deepfake detection using Vision Transformer (ViT) and CLIP-style vision-language model components. It also intends to make detection records immutable through blockchain and distribute results through SMTP, though the available tree does not expose dedicated blockchain or email integration modules.

Primary execution and validation entry points are:

• main.py

• validate.py

Runtime Configuration

config.py is the central likely source of model, dataset, and runtime parameters consumed by the execution pipeline. Python is the primary runtime, with dependencies declared in:

• requirements.txt

• config.py

Detection Pipeline

The main flow is expected to accept image/video-derived visual input, preprocess it into model-ready tensors, execute deepfake classification, and report a result.

Core stages are represented by:

• Input transformation: preprocess.py

• Main orchestration: main.py

• Evaluation/inference validation: validate.py

The boundary between raw media handling and learned inference is preprocess.py, which should produce normalized inputs compatible with the vision models.

Vision Transformer Backbone

The repository contains a local ViT implementation rather than relying solely on an external model package. The backbone and its supporting mechanisms are separated into:

• Transformer model definition: vision_transformer.py

• Shared transformer utilities: vision_transformer_utils.py

• Supporting model pieces: vision_transformer_misc.py

This layer is the primary learned visual-feature extraction boundary. It converts preprocessed image patches into contextual token or embedding representations used by detection logic.

CLIP Model Layer

CLIP-related modules provide an additional multimodal or pretrained embedding path:

• CLIP interface/loading logic: clip.py

• CLIP model definitions: clip_models.py

These files form a distinct model boundary from the native ViT implementation. main.py can select or combine these representations for deepfake classification, while preprocessing must satisfy each model’s expected image format.

Region-Aware Analysis

region_awareness.py indicates a specialized stage that focuses inference on image regions rather than treating every image uniformly. It likely consumes visual embeddings or feature maps from the ViT/CLIP layers and produces region-sensitive signals that improve detection of localized GAN artifacts.

Relevant paths:

• region_awareness.py

• vision_transformer.py

• clip_models.py

Result Handling and External-Service Intent

The README specifies two post-detection concerns: immutable blockchain records and SMTP result delivery. No explicit service adapters, smart-contract code, credentials modules, or mail clients appear in the tree. Therefore, these integrations are either incomplete, embedded in orchestration code, or expected to be supplied externally.

The likely control boundary is:

• Model result generation in main.py

• Result checking/reporting in validate.py

Any production integration should keep blockchain persistence and SMTP delivery outside the model modules so inference remains usable without external network dependencies.