Open source

Proof of Capture is an open-source camera that signs photos into the pixels, and Apple's Reference Image writes its signature into a separate file instead

September 11, 2026 at 2:30 AM PT

The Proof of Capture camera, a 3D-printed enclosure holding a Raspberry Pi Zero, a display board and a shutter button

Image: María Benavente

Why it mattersA team building tooling around image authenticity now has an open reference implementation to read against Apple's closed one, and one that is designed to survive the metadata stripping every social platform does on upload.

María Benavente published Proof of Capture on 10 September 2026, a post describing the open-source camera she and Alex Hornstein built at the Recurse Center this summer, and comparing it to the Apple Reference Image system Apple announced the day before. The camera signs a perceptual hash of the photo into the photo's own pixels, so the signature survives compression and resizing.

The Hacker News thread for the post is at 100 points at the time of writing.

What is on the board

The device, Benavente writes, is a Raspberry Pi Zero, a display board, an ATECC608 cryptographic chip, a shutter button and a 3D-printed enclosure. The ATECC608 holds a public key for verifying and a private key for signing, and Benavente says the private key never leaves the chip once minted, not even for the owner. Tampering with the chip locks it.

The build costs under 100 US dollars, according to the post, and the project source is open.

How the signature survives a share

Benavente writes that the first version hid an exact SHA-256 hash in the last bit of each pixel of the photo, and any JPEG recompression destroyed it. The current version signs a perceptual hash instead, which is a hash of what the photo looks like rather than of its exact bytes, and spreads that signature across the whole image as a frequency-domain watermark using a discrete wavelet transform followed by a discrete cosine transform. The post says the watermark survives WhatsApp-grade compression and still detects content edits.

The reason to hide the signature in the pixels rather than in the file's metadata, Benavente writes, is that the moment a user shares an image, the EXIF is stripped for privacy reasons, and any signature stored there is gone with it.

Where Apple went instead

Benavente describes Apple Reference Image as similar in principle: the sensor signs every pixel at capture time. But Apple, she writes, does not embed the signature into the image itself; the Private Cloud Compute service develops it into a "digital negative" that lives next to the photo, and the verification flow is neither public nor clear yet. Benavente also says Apple has not used the existing open standard C2PA, which is what Nikon, Sony, Leica and Adobe already sign with, and that even though Apple is opening APIs so platforms can verify natively, the root of trust stays inside Apple's Private Cloud Compute.

Benavente is honest about the limits of the approach and Apple's alike: a screen attack still works, she writes, in which someone photographs a screen displaying an AI image and gets a signed photo of a fake. Neither Proof of Capture nor Apple Reference Image nor C2PA fully solves the problem.

Source

Source: María Benavente

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

A Datalog engine for agent memory reports 45 times fewer tokens per answer than full context

Lemmalog is a five-day-old Rust project that treats an agent's memory as a Datalog database with proof trees, and reports 45 times fewer tokens per question on LongMemEval than the full-context baseline.

Source: GitHubOpen source

Apple ships a Reference Image API in iOS 27, iPadOS 27 and macOS 27 for signed photos

Apple's new iPhone 18 Pro signs raw camera sensor data at capture, produces an unalterable reference image through Private Cloud Compute, and exposes a Reference Image API to third-party apps on iOS 27, iPadOS 27 and macOS 27.

Source: PressDev tools

Anthropic launched a public tool to check whether a file was made with Claude, using C2PA credentials and text watermarks

Anthropic put up a web page where anyone can upload an image, audio or video file up to 100 MB and see whether it carries a Claude-signed C2PA content credential, with a separate text-watermark detection API in private preview.

Source: Hacker NewsModels & agents