QuadPath-AE: Authenticated Image Encryption with Keyed Quad-Path Scrambling, Evaluated against Statistical and Deep-Learning Attacks
Images increasingly travel over unreliable channels and are analyzed by learning systems. This raises two questions: can an image cipher be both provably secure and robust to channel damage, and can machine learning extract knowledge from its cipher images? We show that QUAD-PATH, our earlier zig-zag/rotation image cipher, is a key-free, histogram-preserving permutation, and propose QuadPath-AE, a nonce-based authenticated image cipher that turns the quad-path idea into a keyed scrambler. Tiles are traversed along zig-zag, snake, spiral or Hilbert paths under one of eight rotations or reflections, a golden-ratio interleaver disperses neighbor
Resource information
- Format
- zip
- Original repository
- 23723205
- DOI
- 10.5281/zenodo.23723205
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- cc-by-4.0
- Last record update
- Nov 5, 2026
About this resource
Images increasingly travel over unreliable channels and are analyzed by learning systems. This raises two questions: can an image cipher be both provably secure and robust to channel damage, and can machine learning extract knowledge from its cipher images? We show that QUAD-PATH, our earlier zig-zag/rotation image cipher, is a key-free, histogram-preserving permutation, and propose QuadPath-AE, a nonce-based authenticated image cipher that turns the quad-path idea into a keyed scrambler. Tiles are traversed along zig-zag, snake, spiral or Hilbert paths under one of eight rotations or reflections, a golden-ratio interleaver disperses neighbors, and a three-pass diffusion with keyed S-boxes yields full avalanche. Encrypt-then-MAC tile tags localize damage, which is then repaired by inpainting. Confidentiality and integrity are proved under standard pseudorandom-function assumptions. On 39 test images, the cipher passes the statistical tests at the rate of a random source. In a deep-learning evaluation, a convolutional network identifies the source image of QUAD-PATH cipher patches with 99.8% accuracy, and a neural attacker recovers AES-CTR plain images under nonce reuse at 30.19 dB. Against QuadPath-AE, all three neural attacks remain at chance level. A cipher image with one quarter occluded is recovered at 35.11 dB, against 22.40 dB for AES-CTR.