My Steganography

Secure your art, unmask the bots: A mobile workspace that seals invisible pixel signatures, featuring a dual-scanner to extract human keys or identify synthetic AI models.

Comment

##THE STORY OF "MY STEGANOGRAPHY"

Inspiration:

A few weeks ago, a close friend sent me a stunning digital painting, proudly claiming they had spent hours drawing it. As a software developer who closely tracks the AI landscape, my eyes immediately spotted the subtle structural tells of generative AI. I called them out, and we had a laugh, but then, a deeper, alarming realization hit me.

If an artist can so easily fool someone who understands technology, what happens to the general public? People who do not live in the tech loop cannot tell the difference. Furthermore, actual human creators are facing a double-edged crisis:

Automated AI scraping bots are harvesting original art assets off the internet without consent.

When human artists post highly detailed work, they are often falsely accused of "cheating" with AI because they have no way to prove their manual craft.

I built My Steganography to solve both sides of this equation: protecting human intellectual property with encrypted signatures and instantly unmasking synthetic generator structures.

What it does:

"My Steganography" is a dark-cyberpunk creator workspace designed to serve as an invisible cryptographic shield for digital artwork.

The Canvas & Camera Workspaces: Creators can draw original sketches on my integrated canvas or snap physical photos using a live secure camera feed.

The Seal Engine (Steganography): Before sharing their art, creators embed their hand-drawn signature or text credentials. The app injects this data directly into the raw pixel matrix using Discrete Cosine Transform (DCT) manipulation.

The Dual-Engine Scanner: When an image is scanned, the app runs a dual analysis loop:

Path A (Human Verified): If a hidden cryptographic signature is found, the app extracts it and reveals my Blueprint X-Ray View, projecting the vector signature back onto the image.

Path B (Synthetic Detected): If no pixel key is found, my custom-trained Machine Learning model analyzes the image to classify whether it was generated by AI, identifying the specific engine (such as Midjourney, ChatGPT, or Gemini) with 96% accuracy, even if the user uploaded a screenshot!

How we built it:

My architecture is split into a sleek frontend mobile design and a robust local machine learning training pipeline.

The Mobile Client

Built using React Native and Expo, styled to match premium 3D design aesthetics with fluid, high-frequency layout transitions using react-native-reanimated. I integrated a canvas interface using custom vector path listeners and implemented a responsive camera viewfinder overlay.

The Steganographic Core

Instead of relying on fragile metadata (which is instantly wiped out when an image is screenshotted or posted to social media), my app embeds the signature directly into the frequency domain of the image.

Let the original cover image be $I$, and the signature watermark payload be $S$. The stego-image $I'$ is generated by embedding the signature coefficients into selected medium-frequency DCT blocks:

$$I' = \text{Embed}(I, S)$$

When scanned, the extraction algorithm recovers the signature coordinates:

$$\text{Extract}(I') = S$$

The AI Detection Classifier

Most mainstream AI detectors rely on direct download files. My app is built to detect screenshots of AI art. I built an $8$-layer Convolutional Neural Network (CNN) using PyTorch and trained it locally on datasets from multiple generative models (including ChatGPT, Gemini, Midjourney, and Meta AI).

When a screenshot is scanned, the CNN bypasses surface-level metadata and analyzes spatial frequency anomalies. My model computes the conditional probability of the source generator $M$:

$$P(\text{AI} = M \mid I') \ge 96\%$$

Challenges we ran into

The Screenshot Problem: Screenshots compress and distort pixels, which easily destroys simple Least Significant Bit (LSB) steganography. I had to spend nights tweaking the stego-engine to ensure the signature coordinates remained readable even after a screen capture.

Storage and Compilation Limits: As an independent developer working on a laptop with limited local storage, running heavy local mobile emulators was a massive hurdle. I had to pivot to lightweight web simulators and test my Expo bundles directly on real physical testbeds over local networks to keep my development system running.

Small Dataset Overfitting: Training a highly accurate PyTorch model with a small initial dataset of test screenshots was difficult. I had to implement strict data augmentation techniques (rotations, noise injection, and scaling) to ensure the CNN didn't overfit and could generalize to brand-new, unseen images.

Accomplishments that we're proud of

I successfully designed a stunning, cohesive dark-cyberpunk user interface that makes cybersecurity feel accessible, empowering, and modern for visual artists.

I built an AI classifier that doesn't just guess "AI or Not," but successfully isolates the specific model architecture with a confidence level of $96\%$.

I proved that a solo developer, handling both complex machine learning training and fluid frontend mobile UI design, can build a functioning proof-of-concept ready for the global stage.

What we learned

Deep Neural Networks: I learned how convolutional layers map spatial frequencies to classify synthetic textures that the human eye cannot see.

Mobile Image Processing: I gained deep insights into how mobile platforms compress images and how to protect pixel data from being ruined during the save state.

The Power of Purpose: Building with a clear mission—protecting struggling digital creators, kept my momentum alive through every terminal error.

What's next for My Steganography

Gemini API Expansion: I plan to integrate the Google Gemini API to analyze uploaded images and generate automated "Privacy and Scraping Risk Reports" for creators before they post online.

True Decentralized Keys: I want to connect the pixel signatures to Web3 developer identities, ensuring that once a human signature is verified, it is anchored to an immutable creator ledger.

App Store Launch: Deploying the fully bridged PyTorch mobile model to the Google Play Store so real-world digital artists can start shielding their artwork today!

Share this project:

Updates