Tile-Wise Authenticated Selective Encryption and Public-Geometry Leakage Evaluation for Graphite Ore Images

Images used in intelligent graphite ore sorting can reveal grade-related visual and spatial information. Selective encryption reduces the cost of full-image protection and preserves nonsensitive background regions, but the number, location, and shape of protected regions may become public side channels. This paper proposes TASR-AE (Tile-wise Authenticated Selective ROI Encryption), a tile-wise authenticated selective ROI (region of interest) encryption method for graphite ore images. Low-, mid-, and high-grade ore instances are merged into a grade-agnostic ore ROI, dilated for boundary tolerance, quantized onto a fixed grid, and encrypted tile by tile with HKDF (HMAC-based Extract-and-Expand Key Derivation Function)-separated ChaCha20-Poly1305 (an authenticated encryption algorithm). An authenticated sidecar binds tile coordinates, anonymized object identity, algorithm metadata, and the public background. We evaluate correctness, overhead, automatic ROI localization, public-geometry grade-membership inference, decoy-budget trade-offs, ciphertext-position leakage, and adaptive real-ROI recovery. On an audited 121-image graphite test split, TASR-AE with ground-truth masks achieves exact recovery and background authentication while protecting 10.81% of pixels without decoys. Grade-agnostic ROI selection removes deterministic high-grade existence leakage, but full public geometry still yields a High-membership ROC-AUC (area under the receiver operating characteristic curve) of 0.7539. A frozen 576-tile decoy budget lowers the validation ROC-AUC to 0.5088, but the test ROC-AUC rises to 0.6491 while protecting 80.86% of the image. Ciphertext texture does not distinguish real and decoy tiles (pooled ROC-AUC 0.5064), yet a position-and-occupancy adaptive attack recovers real ROI membership with ROC-AUC 0.9868 and mean tile IoU (intersection over union) 0.7051. The results show that TASR-AE provides authenticated content protection with public background visibility, but fixed-budget decoys do not provide strong ROI-location hiding.