Forensic Detection of Generated MRI Imagery Using Autoregressive Modeling and Frequency Analysis
Conference
Mahara, A, Rishe, N, Adjouadi, M. (2026). Forensic Detection of Generated MRI Imagery Using Autoregressive Modeling and Frequency Analysis
. 222-231. 10.1109/WACVW68408.2026.00028
Mahara, A, Rishe, N, Adjouadi, M. (2026). Forensic Detection of Generated MRI Imagery Using Autoregressive Modeling and Frequency Analysis
. 222-231. 10.1109/WACVW68408.2026.00028
Magnetic Resonance Imaging (MRI) is a widely adopted technology for acquiring detailed internal images of the human body for diagnostic purposes. Given its sensitive nature and diagnostic importance, the integrity of MRI imagery must be preserved in both clinical and research settings. However, the rapid advancement of generative AI technologies poses risks of adversarial manipulation, potentially compromising diagnostic accuracy and jeopardizing the entire medical imaging pipeline. To address this emerging threat, we present a systematic investigation of state-of-the-art generative methods, including Generative Adversarial Networks (GANs), diffusion models, and autoregressive models, on MRI imagery, providing a comparative analysis of their generative performance. We introduce MRI-Forensics, a curated benchmark dataset, and show that generative manipulations leave distinct and quantifiable signatures in the frequency domain. With this assertion, we develop a new forensic detection framework that combines Autoregressive Image Models and Discrete Wavelet Transform (AIM-DWT) analysis that reliably detects AI-generated manipulations. By integrating frequency-based decomposition with autoregressive visual modeling, we demonstrate that AIM-DWT effectively extracts unique generative fingerprints from synthesized MRI images. In extensive evaluations using MRI-Forensics and an independent brain MRI dataset, experimental results demonstrate the efficacy of our approach, highlighting its potential to ensure diagnostic accuracy and patient trust in medical imaging in the era of generative AI. The code, pretrained weights for reproducibility, and the MRI-Forensics dataset are available at https://github.com/amaha7984/AIM-DWT.
