Media Safety via Veil, Veil8, VeilGarde

Part 1 of 5

A safe image generation and distribution system requires checks at each generation step and phase against copyrighted works, faces and property, before anything reaches an audience. I started off with detection, with prior research already done on watermarking visual works. Together, Veil, Veil8 and VeilGarde form a trifecta designed to remove most of the risk of serving generative media. The aim is an industry that is tamper-evident, where misuse can be traced and narrowed to a small set of deliberate actors that enforcement can act on effectively. Without these systems, generative image models, open source or closed, still have gaps that have yet to close before they feel secure in an environment moving at a pace that is difficult to keep up with. And I deeply believe we are headed for a world that will treat an absence of provenance credentials within media as a risk signal.

"When an artist takes a photo, they either have a gut feeling or are prioritizing a complex matrix of compositions. Either way, it's a one-of-a-kind shot. Millions of those melted together may look like a bridge, but will never be the same." — A.I. Safety & Deepfakes

Veil

Veil is a pipeline that works with a Thread database, which is a part of Ambient Intelligence Index. Images gathered while using an application like Ambient that are verified to be authentically yours, and images you manually upload via Veil, get stored in a Veil thread. The storage system stores images in a Veil8 encoding, which lets your database hold a translated form of the image that can be understood without being visually recreated, protecting the privacy of one's visual life while enough meaning survives to check for likeness in generations. Veil itself is a harness, a system built to work with any image generation architecture, using Veil8 to encode generations at each step, retrieving results from a Thread database and halting generation if a protected person or identity is detected. Veil is also a pipeline built to make it easy to run image generation models locally or in your own hosting solution. It is designed to be an image generation toolkit with provenance signing, tracking, and protection built in. It is meant to make following policy and managing risk automatic, without having to work through every detail: a safety standard that comes with an easy install, so your users are protected from the start.

VeilGarde

VeilGarde is the provenance signing and tracking system for when a generation passes the halt gate. A generation provided by your system gets signed with a Veil watermark. It's similar to SynthID, but I am designing it to be more robust. A user should be able to upload a photo and check whether it has a Veil watermark, to prove whether an image is real or generated. There are two forms of this watermark: one that signs real images to prove authenticity, and another that signs generations to prove a generation was made with Veil specifically. The second proves the generation came from a system that already had likeness protection, and that it is policy-approved as well: checked against copyrighted works, and royalty-free.

VeilGarde also has a detection mechanism, built to analyze whether a real camera or a generative model produced an image without scanning for watermark artifacts. Its reliability is still growing, and it needs to be kept constantly up to date with the generative media styles found in the wild, but it is available today to forensically analyze any image one would require an autopsy of.

The primary concern is building a policy-validated image generation pipeline, Veil, that is compatible with every generative model offered in the open source world, or integrates directly with enterprise platforms such as Midjourney, while offering two watermarking technologies to create a verified reality. The Veil watermark is in provisional patent status right now, for signing audio and visual works through a custom method that was actually derived from the deepfake detection research. Solving extremely hard problems has its boons at times, discovering other use cases that help in ways we didn't expect. Positioning deepfake detection within the constraints of model-agnosticism, rather than detecting model generation fingerprints from the start, enabled this way of thinking. A common engineering practice of mine is to create hard conditions on purpose. Because it is within constraints that we truly innovate.

Veil8

Back to the ingestion phase: visual media gets stored in a Thread database, whether a personal one or a curated one in the cloud meant to protect at scale, a likeness registry of sorts. The issue with data storage is that it carries privacy concerns and liability with it. When you store the creations or visual identity of others on a hard drive, it is quite sensitive and one is liable for that. I mean, I generally forget to clean up my iCloud photos even...

As an iOS developer who specialized in computer vision and built only for Apple platforms, I treated my iCloud as a permanent testing surface, and what I learned along the way led me to the privacy-preserving ideas now in Veil8. Building camera interactions where a machine can "see something" and then "do something" has always felt like magic to me, but getting to the point where you can iterate on that kind of product is hard and tedious. Training a model means sourcing data, training, tuning parameters and adjusting them run after run until accuracy reaches its target. Sometimes the challenge is how much data you need, but more often it's finding the right amount, with the exact features the model needs.

I prefer a burst approach: scrape at scale, then trim down to a fine-grained dataset. Doing that safely takes care, because a large-scale scraper paging through Google results for matching images can end up pulling from Instagram, X and similar sites if they are allowed by mistake. Without the resources of a company like Google, I've learned to build scrappy systems that make competitive on-device models possible.

Veil8 stores just the features we need to achieve a similarity search. A picture is a visual representation through this language of red, green and blue numbers. A machine doesn't see what we see; a machine sees a list of tuples, and sometimes the alpha channel with it. We see stories. Thankfully, in the field of computer vision we just need the numbers, so Veil8 encodes images into a format a computer understands, not a human, preserving privacy on a human-to-human translation level. It is a method only Veil's pipeline understands, used to find regions of likeness, report attribution concerns, and halt generation early, even protecting individuals from deepfakes.

A Closed Loop

The pipeline, the storage that reduces risk and maintains the privacy of the internet's consumers, and a tool to detect and sign reality together form a closed-loop system for building computer vision solutions in this new age of artificial intelligence. Builders can focus on great products without carrying unmanaged liability, and consumers can trust that their rights are respected. I call that balance, and it is a solid foundation for innovating in this field and pushing the industry further. I firmly believe open source communities are where true innovation happens, where the crazy ones live. What they need is the right infrastructure: systems that make it safe to build, to experiment, and to challenge the status quo. That has always been Silicon Valley at its best... the open one.