Steganography in procedurally generated normal maps via angular half-bin relabeling embedding

Abstract

This paper introduces Half-Bin Relabeling Embedding (HBRE), a novel audio steganography scheme that conceals data payloads within the angular phase of procedurally generated tangent-space normal maps—synthetic assets constructed from tileable fractional Brownian motion (fBm) Perlin noise, eliminating dependency on real-world cover imagery. Unlike conventional spatial domain methods that embed in real photographs, our synthetic carrier is generated without any real-world cover image, giving the encoder full control over carrier statistics and eliminating cover-source bias. The HBRE algorithm encodes one bit per eligible pixel by reassigning only its angular half-bin membership while preserving the within-half phase offset, yielding a near-uniform intra-bin distribution statistically indistinguishable from the unmodified carrier. The system integrates HBRE with AES-256-GCM encryption, PBKDF2-SHA256 key derivation, and Reed-Solomon error correction, with per-pixel adaptive safety margins guaranteeing bit-perfect BC5 texture compression survival. The HBRE algorithm is formalised with a bit-stability proposition, validated empirically, and positioned against Quantization Index Modulation(QIM) and Angle QIM, demonstrating that offset-preserving relabeling strategy and adaptive per-pixel safety margins are the key properties responsible for improved ML resistance. Empirical evaluation demonstrates that HBRE reduces machine learning detection to $$0.270\!\pm \!0.089$$ Area Under the ROC Curve (AUC) under Gradient Boosted Machine (GBM) classification and $$0.106\!\pm \!0.033$$ under Random Forest (RF) classification. Across 165 automated test cases the system achieves: PSNR $$=$$ 39.4 dB at full capacity (approximately 22 KB on a $$512\times 512$$ carrier); Bits Per Eligible Pixel (BPEP) $$=$$ 1.000; AUC < 0.18 against all domain-agnostic detectors; and bit-perfect audio recovery across all nine tested audio formats. Experimental results further confirm robust image fidelity with MSE 13.95, PSNR $$=$$ 36.68 dB, SSIM $$=$$ 0.9751, and ORB similarity $$=$$ 0.8540 on standard image quality benchmarks.