Open-source "proof of capture" camera signs real photos with steganography, not metadata

A day after Apple unveiled Apple Reference Image—its scheme for cryptographically proving a photo came from a real camera rather than a generator—developer Mery Benavente and Alex Hornstein detailed a homemade alternative built at the Recurse Center. Their premise is that detecting fakes after the fact is a losing game, since every better detector just trains the next generator. Instead of authenticating fakes, they authenticate reality at the moment of capture.

The technical trick is where the signature lives. Rather than stashing it in EXIF metadata—which gets stripped the instant an image is shared—they embed a signed perceptual hash directly into the pixels using steganography. An early version hid an exact SHA-256 in each pixel’s last bit, but any JPEG recompression wiped it out; the current build spreads a signed pHash across the image as a frequency-domain watermark (DWT + DCT) that survives WhatsApp-grade compression while still flagging content edits. Signing runs on a self-soldered ATECC608 crypto chip whose private key never leaves the hardware and which locks itself if tampered with. The whole rig is open source and buildable for under $100.

The author contrasts this with Apple’s approach, where the sensor signs pixels but the signature becomes a separate “digital negative” processed through Private Cloud Compute, with a verification flow that remains opaque. The sharper criticism is Apple’s choice to build a proprietary root of trust instead of adopting C2PA, the open standard already used by Nikon, Sony, Leica, and Adobe. None of these systems is airtight—a “screen attack,” photographing a display showing an AI image, still yields a validly signed fake—but the piece frames big-vendor interest in capture-time provenance as a welcome, if imperfect, step.