When you upload a file to a major cloud storage service, something happens before that file settles quietly into a storage bucket. A content scanning system runs a check against it — not a human reading it, but an automated hash comparison against databases of known illegal content. Most services use some form of this. Most users have never been told.
Understanding what cloud content scanning is, how it works, and where it breaks down is relevant to anyone who stores personal files online. The technology exists to catch serious crimes, and it mostly does that. But the systems are imperfect, the transparency is low, and recent research has surfaced privacy vulnerabilities that complicate the picture significantly.
What Cloud Content Scanning Is
Content scanning in cloud services primarily serves one purpose: detecting known child sexual abuse material (CSAM). Services including Microsoft OneDrive, Google Drive, Dropbox, and iCloud scan uploaded files — particularly images and videos — against hash databases maintained by organizations like the National Center for Missing and Exploited Children (NCMEC) and the Internet Watch Foundation.
The dominant technology used is called perceptual hashing. The most widely deployed implementation is Microsoft’s PhotoDNA, which generates a numerical fingerprint for each image that survives common modifications: resizing, color adjustments, cropping. The fingerprint of a known illegal image is stored in a database. When a new image is uploaded to a service, its fingerprint is calculated and compared against the database. A match triggers an alert.
PhotoDNA has been deployed since 2009 and has resulted in tens of thousands of reports of known illegal images. Major platforms using it include Facebook, Twitter (now X), Google, and Microsoft’s own services. It is genuinely effective at detecting known illegal content.
The key word is “known.” Perceptual hashing identifies previously documented images by comparing fingerprints. It cannot detect new illegal content that hasn’t already been documented and added to the database. It is a matching system, not a comprehension system — it does not analyze content or understand what an image depicts.
What Gets Scanned
The scope of scanning varies by provider, but the pattern is consistent: image and video files are the primary targets. Documents, audio files, and other file types may or may not be subject to scanning depending on the service.
Most providers do not disclose which specific categories of files they scan, and the transparency reports they publish describe the volume of illegal content detected and removed without specifying exactly how scanning is implemented or what its technical parameters are. Terms of service for virtually every major cloud platform reserve the right to scan stored content for illegal material.
This is worth noting: when a cloud service scans your files against an external database, your file — or a derived representation of it — is being processed by a system whose full behavior is not visible to you, and the results of which can be reported to law enforcement.
The Vulnerability Research
Recent security research has surfaced serious concerns about perceptual hashing systems that complicate their previously assumed-to-be-benign privacy footprint.
A research paper published through the PseudoDNA project demonstrated two significant attack categories against systems using perceptual hashing, including implementations similar to PhotoDNA.
False positives through adversarial manipulation. The researchers showed that an attacker could construct a benign image — one with no illegal content — that produces a hash matching a known illegal image. This means a person storing innocent personal photos could have a file flagged as illegal content because someone with adversarial intent crafted an image that collides with the hash of known material. A match by a perceptual hash system doesn’t verify illegal content; it verifies hash proximity.
Partial reconstruction from hashes. The researchers also demonstrated that hash values can be partially reversed to reconstruct approximate visual information about the source image. The hashes were designed to be one-way — something like a fingerprint that reveals nothing about the original — but the research showed this property does not fully hold. Hash values, if accessed, can leak information about image content.
CyberInsider has reported on this research specifically, noting that the ability to reconstruct partial visual information from hashes undermines the privacy-neutral framing that has accompanied these systems. A perceptual hash of a personal photo is no longer cleanly separable from the photo itself.
Why This Matters for Personal File Storage
The practical implications for ordinary users are threefold.
You are not informed. Services perform this scanning under terms of service provisions that most users never read in full. There is no opt-out, no notification when your files are scanned, and no way to verify which of your files have been processed against which databases. The scanning is framed as a law enforcement aid and a platform safety measure, which it is — but it is also a form of automated examination of your files.
False positive risk exists. The research on adversarial manipulation means that the detection systems are not perfectly reliable. A person storing photos of their children could, in a worst-case scenario involving adversarially crafted images, experience a false match and the consequences that follow from it — which can include account suspension and law enforcement contact before any human review establishes the error. These cases are rare. They have happened.
Hashes are not value-neutral metadata. The common framing of perceptual hashing — “we don’t look at your images, we only compare a number” — is accurate as a technical description of normal operation but incomplete as a privacy guarantee. The hash itself carries information about image content. If hashes were accessed or leaked, they would not be purely opaque representations.
The Broader Scanning Question
CSAM detection is the most clearly established use case for cloud content scanning, but it is not the only one. Platforms also scan for copyright-infringing content (using audio fingerprinting on video platforms), for terrorist and violent extremist material (using hash-sharing consortiums), and increasingly for other categories of material.
The EU’s proposed Chat Control regulation, which would have required scanning of private messages, was rejected in part because of concerns about exactly this scaling problem: a scanning infrastructure built for one purpose can be expanded to others with fewer public safeguards. The legal and technical apparatus is the same; what it detects is a policy decision.
Cloud scanning for CSAM is legally required under US law for services that discover it — the PROTECT Our Children Act requires that discovered CSAM be reported to NCMEC. Services are not required to scan; they scan voluntarily and under commercial pressure to detect illegal content before regulators find it on their platforms.
What This Means When You Choose a Storage Service
The fact that cloud content scanning exists is not a reason to stop using cloud storage. CSAM detection serves a legitimate and important purpose. But understanding what is happening to your files changes how you think about where to store particularly sensitive personal content.
A service that advertises privacy primarily through encryption should be asked a follow-up question: at what point does encryption apply? If your files are encrypted only after they are scanned — which is the typical architecture, because scanning requires access to the plaintext of the file — then the encryption protects against data breaches but not against examination by the service’s own automated systems before the encryption layer is applied.
Truly private storage, from a technical standpoint, would require client-side encryption before upload: the file is encrypted on your device, and only the encrypted ciphertext ever reaches the server. In this architecture, content scanning cannot occur because the service never has access to the plaintext. The trade-off is that the service also cannot offer AI-powered search or previews, since it cannot read your files.
daftei encrypts files using TLS 1.3 in transit and AES-256 at rest — server-side encryption — which means files are protected against unauthorized third-party access and data breaches. This is the same architecture used by most major cloud providers. It means your files are protected against external attackers but are accessible to server-side processing. The service does not analyze file content for AI organization or training purposes; files are stored and returned to you.
Transparency as a Baseline
The most reasonable expectation for any cloud service is transparency about what scanning occurs. A service that can tell you clearly what it checks, under what legal framework, and what happens when a match is found is operating with better practices than one that buries scanning in terms of service and provides no public documentation.
Most major providers publish some version of this in transparency reports or help center documentation. The specifics of what databases are used, what file types are scanned, and what the false positive process looks like are often harder to find.
If the content you store includes images that you would prefer are not scanned — personal medical photos, images of minors in ordinary family contexts, or any other category of private content — the relevant questions to ask are: does this service scan uploaded images, against what databases, and what is its documented false positive response process?
Those are not paranoid questions. They are reasonable ones, given what the scanning systems actually are and what recent research has shown about their limitations.