CrossGen-MedDeepfake-Detection Weights

Model Sources

• Code repository: link
• Paper: Citation and link will be added once the associated paper becomes publicly available.

Repository Contents

This repository provides the trained Soft Attention and classification-head weights for the architecture-shift and paradigm-shift experiments described in the associated paper. It does not include the frozen SPECTRE encoder.

The following abbreviations are used: R = Real, C = CycleGAN, P = Pix2Pix, and D = DDPM.

Each top-level directory represents one experimental setting and follows the naming convention:

OOD-{Protocol}-{TrainingDomains}vs{TestDomains}

Arch and Paradigm identify the architecture-shift and paradigm-shift protocols. Domains before vs are used for training, whereas those after vs form the test set, which includes real patches and manipulations from generators unseen during training.

Directory	Protocol	Training domains	Test domain(s)
OOD-Arch-RCDvsRP	Architecture shift	RCD	RP
OOD-Arch-RPDvsRC	Architecture shift	RPD	RC
OOD-Paradigm-RCPvsRD	Paradigm shift	RCP	RD
OOD-Paradigm-RCvsRD	Paradigm shift	RC	RD
OOD-Paradigm-RPvsRD	Paradigm shift	RP	RD
OOD-Paradigm-RDvsRCP	Paradigm shift	RD	RCP
OOD-Paradigm-RDvsRC	Paradigm shift	RD	RC
OOD-Paradigm-RDvsRP	Paradigm shift	RD	RP

Each experimental directory contains five independently trained models corresponding to random seeds 42–46. The checkpoints follow this structure:

OOD-{Protocol}-{TrainingDomains}vs{TestDomains}/
└── int_soft_attention_linear/
    └── seed_{Seed}/
        └── training/
            └── checkpoint/
                └── model.pt

For example:

OOD-Arch-RCDvsRP/
└── int_soft_attention_linear/
    └── seed_42/
        └── training/
            └── checkpoint/
                └── model.pt

Each model.pt file contains the trained parameters of the Soft Attention module and binary classification head.

Intended Use

The checkpoints are intended for research on:

• semantic manipulation detection in CT imaging;
• medical image forensics;
• generalization to unseen manipulation generators;
• robustness to generator architecture and generative-paradigm shifts.

Training Data

The released components were trained on spatial and high-frequency embeddings extracted from 3D CT patches derived from the M3DSynth benchmark. M3DSynth contains localized semantic manipulations generated using CycleGAN, Pix2Pix, and DDPM.

The embeddings were extracted using a frozen pretrained SPECTRE encoder. No original CT images or volumes are distributed in this repository.

Usage

The released checkpoints take paired spatial and high-frequency embeddings as input. They cannot be applied directly to raw CT patches.

Use the checkpoints with the following companion resources:

• code repository (link)

The checkpoints cannot be used directly through the Hugging Face inference widget.

Citation

If you use these checkpoints, please cite the associated paper. Citation information will be added after the paper becomes publicly available.