NeuralAntiCheat
PluginMIT

NeuralAntiCheat

This anti-cheat uses neural networks; it automatically detects cheaters and studies their mistakes and kicks. It can be trained. The plugin requires ProtocolLib.

249
Downloads
2
Followers
1 months ago
Updated
📦
3
Versions

📖About NeuralAntiCheat

NeuralAntiCheat
A self-learning, packet-based anticheat plugin for Paper 1.21.1

Overview
NeuralAntiCheat is a modern anticheat plugin built from the ground up for Paper 1.21.1. Unlike traditional anticheats that rely on simple threshold checks, NeuralAntiCheat combines real-time packet interception with a built-in neural network that learns from every player session — getting smarter over time.
The plugin requires ProtocolLib to intercept raw client packets before they are processed by the server. All checks run on a dedicated async thread pool, meaning zero TPS impact on your main server thread regardless of player count.

Core Features

Packet-Based Analysis — Intercepts raw client→server packets via ProtocolLib. Checks analyze actual network data, not just Bukkit events, giving far more accurate detection with less overhead.
Fully Asynchronous — Every check runs on a dedicated NeuralAC-Check thread pool that scales to CPU core count. The main server thread is never touched during check execution.
Lag Compensation — Dynamic ping buffer on Reach checks. Players above 300ms ping are excluded from movement checks entirely. Violation weight is scaled down for laggy connections.
Neural Engine (MLP 16→32→16→2) — A pure-Java multilayer perceptron trained on 800 synthetic sessions. Analyzes 16 behavioral features per player. Score near 1.0 = legitimate, near 0.0 = cheating.
Predictive Movement — FlightCheck and SpeedCheck simulate vanilla Minecraft physics each tick: gravity (0.08/tick), drag (0.98), potion modifiers, surface friction. Flags deviation from predicted trajectory rather than a raw threshold.

Checks (16 total)
Combat
KillAura — detects auto-attack hacks by analyzing attack angles (flags if consistently above 75°), multi-target switching patterns, and inhuman CPS with near-zero variance.
Reach — flags hits beyond 3.05 blocks. Dynamically adds a ping buffer so high-latency players don't get false flagged.
AutoArmor — detects instant 0-tick armor equipping that no human hand can physically perform.
FastBow — flags bows released in under 3 ticks, which is physically impossible with a real bowstring draw.
AntiKnockback — detects near-zero velocity after receiving a hit, meaning the client is silently ignoring knockback packets.Movement
Flight — a predictive physics engine simulates vanilla gravity every single tick. If the player's actual Y position deviates more than 0.5 blocks above the predicted trajectory, it flags. No more false positives from jumps or hills.
Speed — calculates the exact allowed horizontal speed based on sprint state, active potions, and surface type. Flags anything more than 15% above that.
Jesus — flags players standing on water for more than 12 consecutive ticks without sinking, which is impossible in vanilla.
Spider — detects wall-climbing by requiring three conditions simultaneously: upward motion, near-zero horizontal speed (so jumping next to a wall doesn't count), and adjacency to a solid wall for 5+ consecutive ticks.
NoFall — tracks what percentage of falls over 4 blocks actually result in damage. Flags if the player takes damage less than 25% of the time.
Phase — detects Noclip and Phase hacks by checking if the player is inside a solid block for 3 or more consecutive ticks.
Timer — measures how fast movement packets arrive. Flags if packets come in more than 15% faster than the vanilla standard of 50ms per tick.Player
FastPlace — uses a rolling 1-second window. Vanilla holding RMB places about 14–16 blocks per second. The check allows up to 16 and flags at 17+, so normal building never triggers it.
FastBreak — flags more than 25 block breaks per second, which is above the maximum achievable with Efficiency V and Haste II combined.
Scaffold — detects auto-bridging by flagging if more than 70% of block placements happen while moving at high speed without looking down.
InventoryMove — flags movement faster than 0.5 m/s while the inventory screen is open, which is impossible in the vanilla client.

Commands
/nac info nac.admin Plugin status, neural engine state
/nac flags nac.admin View violations and neural score
/nac whitelist nac.adminToggle check bypass
/nac train cheat|legit nac.adminManually correct the neural network
/nac reload nac.adminReload config without restart
/nac alerts nac.alertsToggle alert visibility for yourself

Installation

  1. Place ProtocolLib 5.3.0 in your plugins/ folder
  2. Place NeuralAntiCheat-2.0.0.jar in your plugins/ folder
  3. Start the server — neural engine pre-trains automatically on first launch
  4. Updating from v1.x: delete plugins/NeuralAntiCheat/neural/model.dat before starting

Requires: Paper 1.21.1 · Java 21 · ProtocolLib 5.3.0