PSNR vs. SSIM for Image Steganography Quality Assessment

Core answer

Both PSNR and SSIM estimate imperceptibility by comparing the stego image against the original cover image. PSNR gives a single global error figure and is easy to compute, while SSIM better matches human perception of structural changes. In practice, reviewers and steganographers frequently report both: a higher PSNR indicates the stego image is numerically close to the original, while an SSIM near 1.0 indicates that edge structures, textures and perceived brightness are largely preserved. The main condition affecting their applicability is payload size and embedding position: a small payload spread through an image can leave PSNR and SSIM nearly unchanged, whereas larger payloads concentrated in smooth or highly visible regions can degrade both, often with SSIM responding differently because it is less dominated by any single large pixel difference.

What PSNR measures

Peak Signal-to-Noise Ratio expresses the difference between the cover and stego image as a logarithmic ratio based on mean squared error. In steganography it is calculated from the original image and its modified, payload-bearing version. For a standard 8-bit image, the maximum possible pixel value is 255. Since the logarithm is base 10, the result is reported in decibels.

PSNR = 10 × log10( MaxPixel² / MSE )

MSE = (1 / MN) × Σ [ Cover(m,n) - Stego(m,n) ]²

How to interpret PSNR

• Higher values mean the stego image is closer to the cover image.
• Values around 40 dB or above are generally considered high quality, but interpretation depends on image content, viewing conditions and the application.
• PSNR treats every pixel equally, so it can be overly forgiving of modifications in visually busy areas and overly sensitive to a single strong local error.
• If the images are identical, the mean squared error is zero and PSNR cannot be computed because division by zero is undefined.
• Values below roughly 30 dB usually indicate clearly noticeable distortion, although this threshold is not universal.

What SSIM measures

Structural Similarity Index Measure compares local patterns in luminance, contrast and structure rather than relying only on per-pixel error. In steganography it is applied between the original cover and the stego image. It is commonly calculated over an 11×11 Gaussian-weighted sliding window, producing a local similarity map whose average is the overall SSIM score.

SSIM(x,y) = [ l(x,y) ]ᵅ × [ c(x,y) ]ᵝ × [ s(x,y) ]ᵞ

l(x,y) luminance:  (2μₓμᵧ + C₁) / (μₓ² + μᵧ² + C₁)
c(x,y) contrast:   (2σₓσᵧ + C₂) / (σₓ² + σᵧ² + C₂)
s(x,y) structure:  (σₓᵧ + C₃) / (σₓσᵧ + C₃)

Here μₓ and μᵧ are local means, σₓ and σᵧ are local variances, and σₓᵧ is the local covariance of the two images. C₁, C₂ and C₃ are small stability constants included to avoid instability where local statistics approach zero. When the exponents and constants are merged in the standard simplified form, the result ranges from -1 to 1, with 1 indicating identical structure.

How to interpret SSIM

• A value of 1.0 means the compared images are identical.
• Values above about 0.98 are often treated as nearly invisible modification, though interpretation again depends on image content.
• SSIM is usually less sensitive to uniform brightness or contrast changes that may produce a large PSNR difference.
• SSIM can penalize shifts in edges or texture structure that PSNR may underestimate.
• Extremely small local windows or uniform regions may make local SSIM unstable, so the global average is the commonly reported result.

Key differences for steganography

Property | PSNR | SSIM

Primary comparison | Pixel error and noise power | Luminance, contrast and structure

Range | Scales with dB; no fixed upper bound for images with finite MSE | Usually reported from -1 to 1 after window calculation

Perceptual alignment | Weak: can miss visible structural change | Stronger: closer to human perceptual quality

Sensitivity to one large error | High, because errors are squared | Local windows can limit its effect on the global score

Computational cost | Low | Moderate to high due to sliding-window statistics

Useful role | Fast numerical fidelity check | Perceptual or structural fidelity check

Why PSNR alone is not enough

PSNR is attractive because it is fast and straightforward, but it is a global average. Two stego images with the same PSNR can look very different: one may hide errors in textured areas while another has a visible artifact concentrated in a smooth region. PSNR also assumes that all pixel coordinates are equally important, which does not match human attention to edges, faces and flat surfaces. For this reason, a high PSNR should not be treated as proof that the hidden message is invisible.

Why SSIM alone is not enough

SSIM is more aligned with perception but is not a steganalysis tool. It can remain high even when a small, statistically anomalous modification is present in a few pixels, because the local statistics change only slightly. It is also affected by the selected window size, constants and implementation. SSIM does not detect whether the modification contains a hidden message or whether the chosen embedding algorithm is secure against statistical analysis.

Applying both metrics to stego image evaluation

A practical quality assessment should compare the cover image and the final stego image under the same implementation and image format, including any recompression step. Report PSNR as the overall pixel-fidelity score and SSIM as the structural-fidelity score. If available, inspect a difference map as well: highlighting the largest per-pixel or per-block changes shows where the embedding method is exploiting the image. A method that keeps both PSNR and SSIM high while spreading changes across hard-to-model regions is generally preferable to one that produces a slightly higher PSNR but creates visible or statistically clustered artifacts.

Limitations and caveats

• Both metrics require the original cover image; a stego image checked without its cover can describe compression quality or local sharpness, but not steganographic imperceptibility against the source.
• Neither PSNR nor SSIM proves that the message can be recovered, that the data are confidential, or that statistical steganalysis will fail.
• Steganography is concealment, not encryption; even visually perfect embedding does not make hidden content encrypted or authenticated.
• The metrics should be applied after every processing step that affects the carrier, including palette conversion, scaling, filtering or lossy compression.
• Embedding capacity is not measured by PSNR or SSIM. A high-quality score can coexist with a very small feasible payload, and a lower score can coexist with a larger payload.

Which metric should an implementation report

For a visual-imperceptibility test, report both PSNR and SSIM together. PSNR supplies a reproducible numerical error figure and is useful for comparing methods under the same payload size; SSIM supplements that with structural and perceptual information. Do not replace the quality measurement with a single threshold, because acceptable values depend on image resolution, content, viewing distance and the purpose of the steganographic system.