title: ASR.track1 emoji: 🏆 colorFrom: blue colorTo: indigo sdk: gradio sdk_version: 5.33.0 app_file: app.py pinned: false license: openrail short_description: '[June 2025 Hackathon Submission] [Track 1: MCP Server/Tool]' tag: mcp-server-track

Failed Steganography Detection Tool

June 2025, Huggingface agents/mcp hackathon:

• Track 1 submission

Author: EK

This is an MCP server that is meant to be invoked by an LLM in order to determine if an image has Steganographic Content hidden in it.

• An MCP server is basically a tool that an LLM can invoke or utilize when prompted by the user.

• Steganography is the idea of hiding something (usually a message) inside of an image that is hidden in plain sight. One method for digital Steganography is to manipulate or tamper with the Least Significant Bit of the pixels in the image.

• This page does a good job of explaining Least Significant Bit Steganography and does a good job of visually illustrating steganography in general.

• Detection of Least Significant Bit Steganographic tampering, by itself, does a poor job of determining if an image has anything hidden in it. The code has been changed to hopefully provide a more thorough analysis.

  • Utilises scipy.stats methods: chisquare, skew and kurtosis

• Regular Singular Analysis is when you group pixels and then flip those groupings and then filter them through a some sort of discrimination function that flags groups as being weird or abnormal. This is because when you tamper with an image abnormal "crumbs" are left behind kind of like how you walk in sand and each step leaves an impression just imagine that we are talking in a much less noticable smaller scale.

• Simplified Sample Pair Analysis when you examine neighboring pixels and try to identify abnormal patterns or distributions of even/odd pixel patterns. Untampered images tend to have a predictable pattern and looking for disruptions in this predictability may be an indicator of image tampering.

I have FAILED to producing something that is actually useful, allow me to explain.

TL;DR: My project to detect Steganographic hidden content within images DO NOT produce useful results because any digital image that has been tampered with in ANY way shape or form gets flagged with mixed signals from the various different detection methods used.

Even on a test image that intentionally has a hidden LSB steganographic message inside it will get MIXED flags as some methods will detect abnormality and while others won't pick up on anything. Something like cropping a meme, making changes to the photo's lighting or drawing arrows to emphasize something will throw mixed signals.

The only thing I haven't tested is a brand new crisp photo which I doubt matters because a photo taken from different cell phones or different cameras from different companies probably all have small "microscopic" differences. Which leads me to my next question of what EXACTLY does it even mean for an image to be original?

(You can still upload images and determine which of my rudimentarily implemented detection methods flag your image. But its still not a useful tool.)

Concept and Limitations:

I don't like the idea of AI deep fakes. Like in a few decades what if I told you I could generate video and photographic images of you or anyone committing acts of terrorism or something? The orginal idea was to insert Content Creation into a picture or video upon creation. This is like digitally signing it as your work and then we could hide that signature of ownership inside of the pixels of the image using steganography. So then images or videos will have authentic verifiable author signatures implemented inside of them.

In the past 4chan and reddit and all these people had issues with people reuploading memes for money on youtube or other sites. So people were like what if we added ownership watermarks that are visible on the content? But there are tools for removing watermarks so then it devolved into what if we made the entire content out of water marks so if they removed them then no content. But then it just became annoying and people realized everyone stole memes from each other all the time (Meme Watermark Saga).

• Originally I wanted to create a tool that allowed artists and other content creators to be able to insert Content Credentials or some other related Content Authenticity Initiative related approach like C2PA inside of the image using steganographic tools like lm-steganography
  • Due to time constraints and procrastination I had to alter the project's scope.
  • I also wanted to create a custom gradio component for track 2 to be used by this project also scraped due to above reasons.
  • I now realize that any kind of image alteration would impact detection greatly. A fix for this possibly exists by utilizing machine learning techniques? No clue.
• I wanted to have the embedded credentials tied to the image so that if changes were made to these credentials then visible changes would me made to the image. However this isn't the case with this project.
  • Efforts to mainstream this process would need to happen because it doesn't matter if the tool or implementation is perfect, if no one uses it then none of this work matters.

Further Research:

• In my mind the biggest conceptual obstacle is retaining accuracy of the embedded credentials from user tampering or manipulation. This isn't always malicious in intent, for example, what if you wanted to rescale, crop, or add circles/arrows to emphasize something in the image? These are normal day to day interactions that could affect the integrity of the steganographic embedded message. Also someone who is aware of this process and wants to remove those credentials will probably always find a way around it similar to how the never ending tug of war in cyber security is between threat actors and defenders.

• On the content credentials website they show a time line of certain parts of an image changes over time (a photograph of a penguin in the arctic to an edited image of the same penguin inside of a zoo for example).

  • Another obstacle is how this time line can be maintained? Would it act like a blockchain because many (not all) blockchain authentication processes also require access to the internet. In my mind art being tied to the internet isn't ideal.
  • Also if we are talking about offline creation timelines then wouldn't there need to be a database that could be tampered with?
  • If so then in the future I could see certain artworks behaving like fossils, surfacing with older authentication causing significant shifts in known origins. A situation where for many years it was thought that photo_A was the original rendition of a commonly used meme then seemingly out of a no another rendition even older than photo_A is discovered.
  • If some of these questions are already answer then I apologies in advance for my lack of investigating into the methodology of Content Authenticity Initiative.