A new method embeds encrypted audio in computer-generated texture maps and recovered it exactly across nine tested audio formats.
The method, called Half-Bin Relabeling Embedding, stores audio in the angular information of procedurally generated texture maps made from computer-generated noise. It combines encryption, error correction and per-pixel safety margins designed to help the hidden data survive BC5 texture compression.
At full capacity, a 512-by-512 carrier held about 22 KB of data with a reported PSNR of 39.4 decibels. The researchers also tested how readily machine-learning classifiers could detect the embedded information and compared the method with other data-hiding approaches.
What the texture tests found
The system encoded one bit per eligible pixel and recovered the embedded audio exactly in all nine tested audio formats. Across 165 automated test cases, it reached 1.000 bits per eligible pixel at full capacity, or about 22 KB in a 512-by-512 carrier, with a PSNR of 39.4 decibels.
The reported machine-learning detection results included an area under the receiver operating characteristic curve of 0.270 ± 0.089 for a gradient-boosted machine classifier and 0.106 ± 0.033 for a random-forest classifier. The researchers also report an area under 0.18 against all domain-agnostic detectors. On standard image-quality tests, the system produced a mean squared error of 13.95, a PSNR of 36.68 decibels, a structural similarity score of 0.9751 and an ORB similarity score of 0.8540.
Why synthetic carriers matter
Because the carrier images are generated rather than taken from the real world, the encoder does not depend on a particular photograph or its source characteristics. That gives the system direct control over the texture used to carry the data.
The results indicate that the approach can combine audio hiding with encryption, error correction and resistance to the tested forms of automated detection while retaining visual similarity. Its practical value, however, depends on how it performs beyond the synthetic carriers and test conditions examined here.
Evidence and limits
The evidence comes from a formalized algorithm, a mathematical bit-stability proposition and empirical testing reported in the paper. The study tested 165 automated cases, nine audio formats, synthetic procedurally generated carriers and specific image-quality, compression and machine-learning detection benchmarks.
The abstract does not report testing on ordinary photographs or other real-world cover images, nor does it establish performance against every possible detector or compression system. The reported results therefore describe the tested setup rather than all possible images, audio files or analysis tools.
// Source
Scientific Reports · 2026 · DOI: 10.1038/s41598-026-69716-8
Authors: Porchezhian R, Naman Jain, L. Mary Shamala
Institutions: Vellore Institute of Technology University
