SenseDefend

Synthetic-media instruments, by article and effective date

Why detecting a mark is a measurement

A record stored in a container is either there or not. A signal mixed into content has to be recovered from material that has been re-encoded, and recovery produces a number somebody has to decide about. As of 2026-09-12.

Two ways to be wrong instead of oneA record in a container is present or absent. A signal recovered from re-encoded content produces a score against a threshold, so a system can report marks that are not there and miss marks that are.The detector reports no mark. What follows?The route is known to preserve the signalInformativeAbsence is meaningful, because presencewould have survivedThe route is unknown or lossyNothing followsProcessing moves the score without changingwhether content was generatedUseful for confirming origin, not for refuting it
Fig. 1 A published detection rate without a false positive rate describes half of a trade-off.
How each kind of mark is read, and how it fails. Recorded 2026-09-12.
Kind of markHow reading it can go wrong
A record in the containerIt is absent; there is nothing to misread
A signal in the contentA score near the threshold, either way
A notice in the frameA person does not notice it
No mark at allNothing distinguishes it from a stripped file

Inclusion rule. Failure modes of reading each marking mechanism. Order. From the most binary to the least.

1Two ways to be wrong instead of one

A threshold set low reports marks in unmarked material. Set high it misses marks in heavily processed material. Both errors are real and they cannot both be eliminated by choosing better.

Which of the two a system prefers is a policy decision rather than an engineering one, and it is almost never published. A published detection rate without a false positive rate describes half of a trade-off.

2Processing moves the score, not the truth

Compression, cropping and mixing all reduce the signal without changing whether the content was generated. So a negative result on a heavily handled file says something about the handling rather than about the content.

That is the opposite of how a negative result is usually read. Absence of a detected mark is regularly treated as evidence of authenticity, and it is not evidence of anything at all.

3Why robustness claims name operations

A claim that a mark survives is only meaningful against a named list of operations, because survival is not a property of the mark alone. Cropping, filters, rate changes and compression are the usual list.

What such a list rarely includes is a rate for each operation. Designed to survive is design intent; a figure would be a measurement, and measurements are what nobody publishes.

4What this means for anybody holding a file

A positive detection is informative. A negative one is not, unless the route the file travelled is known to preserve the signal, which almost nothing does reliably.

So the practical use of these systems is confirming origin rather than refuting it. A tool that cannot say a file is clean is still useful; a tool used as if it could is dangerous.

Background on mechanism and practice. Nothing here is attributed to a product, and nothing here is a reading of any instrument. The sourced material is on the generator table. Related: Verify the route, Version fields.