Plain-English explanations of every detection method — written for business owners and compliance officers, not researchers.
Chi-square — JPEG Frequency Manipulation
When you save a photo as JPEG, the image is broken into 8×8 pixel blocks and converted into frequency coefficients. Someone modified the last digit of these coefficients to encode hidden information. The chi-square detector measures how "equalized" the distribution of these digits is — natural images are non-uniform, embedded images approach 50/50. A score above 0.65 indicates embedding.
What to do: If this is your watermark, it's working. If unexpected, someone embedded data in the JPEG frequency domain — a forensics professional can extract the payload.
RS Analysis & SPA — Pixel-Level Data Hiding
Every pixel has a value from 0–255. The last digit (the "least significant bit") has almost no effect on how a color looks. Changing a pixel's value from 128 to 129 is invisible to the eye. RS and SPA measure whether the even/odd distribution of pixel values has been statistically altered — natural images have a specific imbalance that disappears when data is embedded.
What to do: The payload rate estimate tells you what fraction of pixels were modified. Above 30% is significant. This technique is commonly used for covert communication.
EOF Detection — Hidden Data After Image End
Every image file has a defined ending point. JPEG files end with bytes FF D9. Everything after that point is ignored by photo viewers — but it's still in the file. Someone appended data after this end marker. Your photo app shows the image normally and never reveals the hidden content.
What to do: This is the most detectable steganography technique. Preserve the original file as evidence. The hidden bytes count in your report tells you exactly how much data was appended.
EXIF Analysis — Metadata Anomalies
Every photo carries hidden information — date taken, GPS location, camera model, software used. This is called EXIF metadata. Stegmarc checks for: known steganography tool names in the software field, creation/modification time mismatches indicating the image was edited after capture, missing metadata (commonly stripped to hide evidence), and unusually long comment fields used to store hidden text.
What to do: Missing EXIF combined with other detector hits is a strong indicator of deliberate manipulation. Consult a digital forensics professional.
LSB Histogram — Pixel Uniformity Analysis
In natural photos, the last bit of each pixel's color value follows a non-uniform pattern tied to the image content. When data is embedded, this pattern becomes artificially uniform — closer to a perfect 50/50 even/odd split. The LSB histogram measures this uniformity on a 0–1 scale. Natural images score around 0.5. Embedded images approach 0.85+.
What to do: This detector catches embedding methods that evade RS and SPA, including LSB matching. A high uniformity score with no other detectors firing suggests a sophisticated embedding technique.
TrustMark — Adobe C2PA Watermark
TrustMark is Adobe's watermarking system for the C2PA content provenance standard. It embeds an invisible, robust watermark directly into image pixels using a neural network. Unlike visible logos or captions, this watermark survives JPEG compression, resizing, cropping, and color adjustments — and can be decoded to reveal who created the image and when. It's used by AI image generators to label their output for EU AI Act compliance.
What to do: The decoded payload in your report is the identifier the creator embedded. Visit contentauthenticity.org to learn more about verifying content provenance.
Stable Signature — Meta AI Watermark
Stable Signature is Meta AI's system for identifying images generated by Stable Diffusion. Instead of adding a watermark after generation, it modifies the AI model itself to embed a 48-bit identifier into every image it produces automatically. The identifier can be traced back to the specific model version that generated the image — relevant for EU AI Act compliance and copyright disputes.
What to do: A bit accuracy above 80% is a confident detection. The decoded bit pattern serves as the AI provenance identifier for compliance documentation.
SRNet — Adaptive Steganography
Adaptive steganography is sophisticated — it analyzes the image content first, then hides data specifically in the most visually complex areas where changes are hardest to detect. Algorithms like WOW, S-UNIWARD, and HILL are undetectable by classical statistical methods. SRNet is a deep learning model trained to detect these advanced techniques by recognizing the subtle patterns they leave in the image's noise structure.
What to do: This is a sophisticated, technically advanced embedding technique. A stego probability above 70% warrants escalation to a digital forensics team. Preserve the file in its original state.