Deep learning-based quantum-secure framework for authentication and encryption of volumetric NIfTI neuroimaging data

As cloud-based healthcare systems, telemedicine platforms and the Internet of Medical Things (IoMT) become more prevalent, the demand for secure authentication and protection of volumetric medical imaging data has grown significantly. Specifically, images provided by the Neuroimaging Informatics Technology Initiative (NIfTI) contain important information that needs to be safeguarded from unauthorized access, tampering, and new threats from quantum computing. In this paper, we propose a token-mediated authentication and encryption framework that establishes functional dependencies among deep feature authentication, chaotic encryption, and post-quantum token protection. To obtain strong feature representations from multi-modal MRI volumes, a 3D Siamese DenseNet architecture is used and authentication tokens are computed by applying a SHA-3 cryptographic hashing function. For the image encryption, a double pendulum chaotic system is used to generate highly sensitive keystreams to ensure confidentiality, and for protecting authentication tokens, ML-KEM lattice-based post-quantum cryptography is adopted against classical and quantum adversaries. This proposed framework was assessed with the multi-modal MRI dataset from BraTS2020, which consists of FLAIR, T1, T1CE and T2 modalities. Experimental results reveal high level of encryption security with a good NPCR, UACI, entropy and statistical analysis results. The results of the decryption process achieved PSNR values greater than 58 dB and SSIM values very close to 0.9999, which shows high-quality image reconstruction and preservation of diagnostic image quality. The results of authentication experiments demonstrated that the DenseNet model proposed in this work obtained better results than the CNN, ResNet and EfficientNet models. The robustness evaluation revealed that the proposed model was able to maintain its performance even in the presence of noise, cropping, compression, rotation and combined attacks. Moreover, experiments with adversarial tampering succeeded in finding all ROI and metadata alterations, and accurately identifying altered areas. The framework proposed in this study resulted in an average processing time of ~ 234.13 s/volume, exhibiting practical usability for secure storage and transmission of medical images. The experimental results demonstrate that the proposed framework provides effective authentication, secure image protection, and high-fidelity reconstruction for volumetric NIfTI neuroimaging data. An additional independent evaluation on twenty volumetric COVID-19 CT cases further demonstrated the transferability of the encryption and reconstruction components beyond the BraTS2020 MRI dataset. The framework incorporates ML-KEM-based post-quantum protection for authentication-token exchange, while the security of the chaotic encryption component is supported by the experimental analyses presented in this study rather than by a formal post-quantum security proof.