Inspiration

Seeing so many creators and digital artists express frustration over major AI labs scraping online art, designs, and visual work without attribution or consent. We wanted to build a practical defensive shield that gives individual creators agency over their own digital assets.

What it does

SKOMKEL applies adversarial perturbation and encryption envelopes to visual assets (art, illustrations, designs, photos). To human viewers, the images remain clear and high quality; however, when web crawlers or automated AI vision scrapers attempt to ingest or train on the data, the adversarial noise causes feature extraction to collapse into useless, incoherent outputs.

How we built it

We architected the prototype using Python, adversarial noise generation models, and Claude Code for iterative scaffolding. We referenced recent research papers on data poisoning and cloaking defenses, consulting patterns from open-source privacy projects, and deployed the compute-intensive open-weights evaluation workloads on Kaggle GPU instances.

Challenges we ran into

Coordinating automated code agents across complex cryptographic dependencies proved tricky, with frequent context drift during multi-file refactors. Additionally, navigating strict platform API guardrails required us to host our open-weights benchmarking pipeline independently on Kaggle rather than relying on locked commercial endpoints.

Accomplishments that we're proud of

Successfully generating an end-to-end cloaking pipeline where sample artwork retains visual fidelity for human portfolios while completely degrading reconstruction quality on test vision models.

What we learned

We gained hands-on experience with adversarial feature perturbations, open-source model deployment on Kaggle, and the practical challenges of balancing perceptual quality against data poisoning efficacy.

What's next for SKOMKEL

Expanding support for batch processing across full artist portfolios, building a lightweight browser extension for one-click upload cloaking, and testing robustness against adaptive image preprocessing defenses.