Image-in-image steganography hides one image inside another. The cover image carries the visible content; the payload image rides underneath, encoded into the least significant bits or learned embedding dimensions. Modern schemes use neural networks to learn the embedding, producing stego-images visually indistinguishable from their covers. The concealment is excellent by any perceptual metric.
Mallet and Bas (arXiv:2603.11876, 2026) show that the concealment is statistically detectable with an eight-dimensional feature vector and 84.6% accuracy. Classical steganalysis methods push detection past 99%.
The mechanism is ICA — independent component analysis. The embedding process combines two independent image sources (cover and payload) into a single output through a learned linear mixing operation. The mixed signal's wavelet coefficients have different statistical moments (skewness, kurtosis) than those of a natural image, because natural images are single sources and stego-images are mixtures. ICA was designed to solve exactly this problem: given a mixture of independent sources, decompose it. The steganographic embedding creates the precise statistical signature that ICA is built to find.
The structural point is not that this particular detection works. It is that the embedding operation and the detection operation are conjugates. Hiding information by mixing two sources together is the mathematical inverse of finding hidden information by separating a mixture into sources. Any steganographic scheme that operates by combining independent signals creates a signal that source-separation algorithms can detect, because detecting mixtures is what source-separation algorithms do. The defender's tool (mixing) defines the attacker's tool (unmixing).
This is distinct from the watermark-removal problem (where the attacker's generative model operates in the same space as the defender's embedding). Here, the relationship is tighter: the operations are inverses. The embedding is the detection, viewed backward. The concealment scheme does not merely create a weakness that someone might exploit — it creates the exact complementary structure that an existing mathematical framework is designed to find. The act of hiding by mixing is an invitation to anyone who can unmix.
The question this raises: is there a steganographic embedding that does not create a conjugate detector? It would need to produce output that is statistically indistinguishable from a single source — not merely visually indistinguishable. Perceptual metrics test whether humans notice. Statistical tests test whether the data's mathematical structure has changed. The gap between perceptual and statistical indistinguishability is where detection lives.