AI anticheats for Minecraft: what neural networks can and can't do

In short: a neural network is the best tool against combat cheats that pretend to be human, but it's no substitute for physics, building and inventory checks. That's why AI works best together with rules inside one anticheat. That's how DGuard — the best anticheat for a modern Minecraft server — is built: in combat it runs its own AI model, AI Combat, trained on billions of gameplay data points from the most popular PvP servers, while checks and rules cover every other front.

What an AI anticheat is#

A classic anticheat works on rules. The game runs at 20 ticks per second and follows exact physics: a player can't run faster than the game allows, hover without an elytra, or hit beyond reach. A rule checks whether physics was broken — and if it was, that's proof.

An AI anticheat works differently. The plugin records combat features — camera rotations, clicks, hits, the target's position — and hands them to a machine learning model. The model was trained in advance on a large set of labelled fights, where it's known whether a human or a cheat was playing, and it judges which one the player's style looks more like.

The model usually runs in the developer's cloud rather than on your server: it needs retraining, it needs compute, and it's harder for cheat developers to study there. Commercial anticheats use machine learning too — Polar, for example, says on its website that its models analyse player behaviour in real time and that most of its checks run in Polar's cloud.

Where neural networks beat rules#

The nastiest combat cheats don't break physics. AimAssist just pulls the crosshair toward the target, a TriggerBot swings the moment the crosshair is already on an enemy, and a carefully tuned KillAura stays within reach and pauses between hits. Every single hit looks possible — a rule has nothing to grab onto.

What shows is the style over time. A human overshoots and corrects their aim, reacts with varying delays and makes different mistakes. A cheat repeats its own style fight after fight, even when it adds randomness. A model that has seen plenty of human and cheater fights picks up those patterns where a rule only sees legal hits.

• It catches "humanised" cheats that stay within the game's physics.
• It judges style across many fights, not a single hit.
• It learns from new recordings: once a cheat is in the training data, it gets recognised without writing a new rule.

Where neural networks stop#

• They need data. A model only knows what it was trained on. A new cheat that looks nothing like its examples can slip past until recordings of it make it into training.
• They need the cloud. The model runs on the developer's servers: your server needs an internet connection, and combat measurements leave it. Check exactly what gets sent and how long it's kept.
• A probability isn't proof. The model says "looks like a cheat" but not why. That makes it hard to show a banned player what they did, so such bans need a clear breakdown, and disputed cases need a human.
• Strong players look like cheaters. The better the player, the closer their aim gets to "perfect". A model's threshold is always a trade-off between misses and false flags.
• Physics is best caught with physics. For Fly, Speed, Timer, NoFall or Phase the game has an exact answer: that's impossible. A physics check is more precise here than any model, and it doesn't need a cloud.
• Combat isn't the whole server. A combat network doesn't look at Nuker, Scaffold, AutoArmor, AutoFish or impossible packets. On anarchy and survival servers you can't do without checks on those fronts.

AI Combat — the strongest AI model in anticheats#

AI Combat is DGuard's own AI combat model. It was trained on billions of gameplay data points collected while working with and testing on the most popular PvP servers, and today it is the strongest AI model in anticheats. The model runs in DGuard's cloud check together with the aim rules and TriggerBot analysis — there's no separate product to install.

• What the model gets. A recording of the fight — every client tick around hits on players: camera turn, movement relative to a session point, movement keys pressed, hits, clicks and slot changes with their timing, ping, the type of items in hand and what the server did to the player.
• What the model doesn't get. Nicknames, UUIDs, IP addresses, chat, coordinates and world names: players are random session numbers.
• How long it's kept. In memory only: the recording is cut into windows, scored and deleted with the session after 10 minutes of silence.
• What it learns from. Only data-collection archives that the server owner turns on and sends with /dg data; collection is off by default. The details are in section 9 of the privacy policy.
• What it costs. AI Combat comes with DGuard on every plan, the free one for up to 30 players online included.

How the model fits in, which fronts it covers and how the cloud check is turned on: the AI Combat page.

What DGuard does#

DGuard is a multi-front anticheat: 11 checks and 80+ rules watch combat, movement, building and world, inventory and packets at the same time. Physics checks catch Fly, Speed and Timer, placement analysis catches Scaffold and Nuker, inventory analysis catches AutoArmor and XCarry, and packet order within a tick gives away a KillAura that rotates before the hit.

In combat DGuard has its own rules, a combat model trained on real cheaters on a live server with 500+ players online and the AI Combat neural network, and the team closes any new bypass within a couple of days. The aim rules, TriggerBot analysis and AI Combat run in DGuard's cloud check: the plugin sends game measurements — camera rotations, positions relative to a random point with no world coordinates, clicks and hits — under random session numbers. They live in memory only and are deleted after 10 minutes of silence, while names, IPs and chat stay on your server.

The key difference from a black box: DGuard bans on evidence. Every flag is weighted evidence, a ban comes only at 400 points across three separate episodes, and every ban comes with a report breaking the evidence down. You can show it to a moderator, or send it to support if a ban turns out wrong. How DGuard stacks up against other anticheats: The best Minecraft anticheat.

A neural network and rules in one anticheat#

AI is strongest when paired with rules. In DGuard they work together: AI Combat reads a fight tick by tick, the aim rules and TriggerBot analysis check what can be measured exactly, and the movement, packet, inventory, building and world checks cover every other front. A cheat would have to fool all of it at once.

DGuard's cloud check confirms a ban from the evidence dossier your server sends, and you set how strict the punishments are in config.yml: a ban or a kick for each cheat feature and evidence thresholds to fit your game mode.

Does your server need an AI anticheat?#

• PvP, duels, BedWars — yes: aim assist and TriggerBot decide fights here. In DGuard the AI Combat neural network already works in the cloud check, so there's no second anticheat to install. More in Anticheat for a PvP server.
• Anarchy and survival — start with an anticheat that covers every front: cheats hit everywhere here — Nuker and FastBreak in the mines, Fly and Speed across the map, automation on the farms. AI for combat comes next. See anticheat for anarchy.
• A small server — DGuard is free up to 30 players online, and AI Combat works on that plan just as on the paid ones: protection is the same on every plan.

How to judge an anticheat on everything else — versions, load, price and honest claims — is covered in How to choose a Minecraft anticheat: 8 questions.

FAQ#