Apple introduces a system for verifying photo authenticity

التقنيات2026-09-17, 10:19
Apple has announced Apple Reference Image, an opt-in camera mode that produces a "reference" shot with hardware-backed provenance. The system is meant to help distinguish photos genuinely captured by an iPhone's sensor from images generated or altered by AI.
How it differs from C2PA The industry-standard C2PA (Content Credentials) attaches provenance metadata after the shot is taken, which leaves it vulnerable to tampering at any point in the editing chain. Apple instead anchors authenticity at the moment of capture — inside the sensor itself, before the frame is ever accessible to the operating system.
How it works • During manufacturing, the sensor itself generates an ECDSA P-256 key pair and never releases the private key. The Secure Enclave Processor (SEP) creates its own separately attested key, and a factory CA binds the sensor and the SEP into the device manifest. • The device regularly (roughly every 15 minutes) obtains a trusted RFC 3161 timestamp from Apple's timestamp service. • In Reference mode, the sensor reboots into a dedicated mode, digitizes the incoming light and signs the frame together with the timestamp using its private key inside the sensor. The SEP then adds its own signature on top. The result is a "secure digital negative" in DNG format. • "Developing" the shot takes place in Private Cloud Compute (PCC): it verifies the sensor's and the SEP's signatures, confirms they belong to the same device, checks the time bounds, and computes a confidence score — a neural-network assessment of whether the frame has the physical characteristics of a real photograph. • The final signature is produced by a hybrid post-quantum algorithm — a combination of RSA-3072 and ML-DSA-87. It is embedded in the JPEG, and any client can verify the signature and confirm that the photo has not been placed on the revocation list — a list of revoked images whose signatures are no longer considered valid (for example, if a sensor has been compromised).
What is guaranteed • the frame reflects exactly what the sensor captured — the OS cannot tamper with the pixels; • a lower and an upper bound on the time of capture; • signatures that are resistant to quantum attacks; • the ability to revoke both individual photos and all photos from a specific sensor; • privacy: the image content is inaccessible even to Apple (the revocation service stores only metadata — the photo's GUID and sensor identifiers), and the timestamp is requested over Oblivious HTTP (RFC 9458), so the service never learns the IP address.
Limitations • The system confirms the authenticity only of shots taken in Reference mode — it does not verify ordinary photos. • A user cannot directly confirm that PCC processing was performed correctly: one has to trust that every PCC build is recorded in an immutable, tamper-proof transparency log. • The legitimacy assessment relies on a neural network with hidden weights, so it is probabilistic in nature. • For now the mode is available only as an opt-in on the main sensor of the iPhone 18 Pro and iPhone 18 Pro Max.
What it means for researchers From a security standpoint, Apple Reference Image is interesting first and foremost as a new chain of trust that unites a hardware data source, cryptographic attestation and remote processing into a single mechanism. Points of interest include protecting the "sensor — SEP" binding against component transfer or substitution, the integrity of the confidence score, the attestation and processing mechanisms in PCC, the robustness of the timestamp chain, and the revocation process. It is also worth noting that the system can confirm an image's origin and integrity, but not the truthfulness of the event it captured. Apple Reference Image should therefore be seen not as a fix for deepfakes, but as infrastructure for verifying the provenance of specific visual data.
المورّدون
المنتجات
المزيد