In recent months, KamAZ and Ural trucks painted not in the usual khaki but in contrasting black-and-white stripes or wavy “leaf” patterns have been spotted with increasing frequency on roads in occupied territories. At first glance, this seems to expose the enemy. But the purpose of these patterns is different: to confuse the machine-vision systems equipping modern Ukrainian strike drones.
Camouflage for a Neural Network
One of the first to write about the new camouflage used by Russian forces was The Economist, comparing it to “dazzle camouflage” — a technique used by Britain’s Royal Navy during World War I. Back then, ships were covered with geometric patterns to make it harder for the enemy to assess course, speed, and distance for a torpedo attack.
1918. Photograph from the Bureau of Ships Collection at the U.S. National Archives.
The Russians are pursuing a different goal: to push an object “outside the distribution” of the data on which the neural network was trained.
The thing is, any computer-vision model does not work on the principle of human thinking — it does not know what a “truck” is in a general sense. It learns statistical patterns from a vast set of photographs: typical cab contours, body proportions, wheel placement, surface textures, characteristic colors, and the contrast between individual elements.
If during training the model saw millions of ordinary trucks — green, sand-colored, white, or camouflaged — but almost never encountered vehicles painted with broad black-and-white stripes, it may perceive such an object as something entirely different or fail to classify it at all.
This phenomenon is called going “out of distribution” of training data. A neural network works well as long as it sees objects similar to those it was trained on. But when shown an unusual variant of a familiar object, its accuracy can drop sharply. It is precisely this effect that was described by University of Texas at Austin engineer Todd Humphreys.
Similar examples have long been known to researchers. For instance, sticking a few specially selected stickers or strips on a “STOP” road sign is enough for an algorithm to mistakenly identify it as a speed-limit sign or even ignore it altogether.
In another experiment, researchers 3D-printed a small turtle that humans unmistakably recognized as an animal, while the neural network consistently classified it as a rifle due to characteristics of its shape and texture.
The same principle applies to military equipment: an unusual paint scheme does not hide the vehicle from the human eye, but it can confuse an algorithm searching for familiar visual patterns.
What Analysts Say: There Is an Effect, but It’s No Silver Bullet
Geert De Cubber, an autonomous-systems specialist at the Royal Military Academy of Belgium, noted in a comment to RFE/RL:
When such patterns are absent from a drone’s database, they can quite plausibly affect the detector’s performance — since models are trained on strictly limited sets of categories in order to avoid false positives.
And Nick Reynolds of London’s RUSI added an important caveat: the camouflage is almost useless against thermal imagers, as it does not conceal the engine’s heat signature.
Kyiv Independent links the appearance of this camouflage to a sharp rise in Ukrainian medium-range strikes on logistics. In May 2026 alone, 130 hits on Russian transport were recorded, and the depth of strikes on supply routes grew to 200 kilometers from the front line. This is precisely why the “zebra” pattern has become most widespread on routes along highway P-280 — the only land supply artery for fuel to occupied Crimea.
A Mirror Process: Drones Are Being Camouflaged Too
Tellingly, the tactic is not one-sided. According to The Economist, some Russian drones are also receiving their own “zebra” paint scheme — presumably to make it harder for Ukrainian interceptor drones to identify them.
A spokesperson for Brave1 — Ukraine’s government unit for accelerating defense technologies — confirmed to RFE/RL and The Economist that Russians are indeed continuously testing new camouflage variants, but stressed that the Ukrainian side adapts faster.
Brave1 also emphasized that the decision to strike a target is always authorized by a human, and AI only assists at certain stages of the process. This applies in particular to strike drones of the Hornet type, which are actively used in medium-range operations.
A Short-Lived Arsenal
The experts’ key point: the advantage gained from any specific pattern is temporary. As soon as enough “zebra” trucks appear to make it into training datasets, models will begin confidently detecting them as well.
This is essentially a classic arms-race cycle: new camouflage — retraining the classifier — new camouflage. Humphreys states plainly that any given paint scheme has a short shelf life. In this race, the winner is not the one who devises the cleverest pattern, but the one who updates the dataset faster.