NullAntiCheat | ML Kill Aura Detector (CatBoost) + Staff Suspect Menu

[mzxcbaisb]

NullAntiCheat
ML Kill Aura Detector for Paper
Collect → Train → Detect. Built for staff review, not instant auto-bans.

​

Overview
NullAntiCheat is a behavioral kill aura module for Paper servers.
It does not rely on classic hardchecks alone. Instead it:

• Listens to combat packets via PacketEvents
• Extracts aim / reach / motion / reaction features from critical hits
• Lets you label players as legit or cheat while testing
• Trains a CatBoost model offline (Python trainer included)
• Scores live hits and ranks players in a staff suspect GUI

Think of it as a NetVision-style combat intelligence layer:
first you gather real fight data, then the model helps moderators decide.

Important: this is not a full all-in-one anticheat.
It focuses on kill aura / aim combat scoring and staff tooling.
No built-in movement HAC suite, no auto-ban pipeline by default.



How it works

1) Collect
Put the plugin into collect mode and mark testers:

/nac mode collect
/nac track Steve legit
/nac track SoftUser cheat

Only tracked players write combat hits into:

plugins/NullAntiCheat/dataset/hits.jsonl

2) Train
Use the included Python trainer (CatBoost by default, ExtraTrees fallback):

pip install -r tools/requirements.txt
python tools/train_model.py \
  --dataset plugins/NullAntiCheat/dataset/hits.jsonl \
  --output plugins/NullAntiCheat/model/model.json

No Torch/CUDA required.

3) Detect
Load the model and switch to detection:

/nac mode detect
/nac model reload
/nac menu

Or use hybrid to detect and keep collecting labeled data at the same time.



Features

• PacketEvents combat + rotation tracking
• ~100+ behavioral features (aim error, reach margin, windowed yaw/pitch deltas, reaction gaps, CPS, target switches, motion context)
• Labeled dataset collection (legit / cheat)
• Offline CatBoost training pipeline included
• Native JSON model loading inside the plugin
• Live suspicion scoring on critical hits
• Staff suspect menu with ranked player heads
• Configurable suspicion thresholds (strong / medium / low)
• /nac info with UUID, playtime, K/D, ping, IP, location + recent ML scores
• /nac playtime helper
• Modes: collect / hybrid / detect
• Fully configurable messages (Russian defaults included)
• Lightweight Java plugin — ML training stays offline

What you can do with it

• Build your own kill aura dataset on your arena / kit / ping profile
• Train a model that fits your community, not a generic public dump
• Give moderators a suspect list instead of blind bans
• Run controlled tests: clean PvP vs kill aura clients
• Retrain after new cheat styles appear
• Use as a combat intelligence layer next to Grim / Vulcan / etc.

Commands

/nac menu
/nac info <player>
/nac playtime <player>
/nac mode <collect|hybrid|detect>
/nac track <player> <cheat|legit>
/nac track list
/nac tracking stop <player|all>
/nac model reload
/nac dataset flush
/nac reload

Aliases: /nullanticheat, /anticheat



Permissions

nullanticheat.menu
nullanticheat.tracking
nullanticheat.info
nullanticheat.playtime
nullanticheat.reload
nullanticheat.admin

All default to op.



Dependencies

• Required: PacketEvents
• Server: Paper 1.16.5+ (developed around Paper API 1.16)
• Training machine: Python 3 + packages from tools/requirements.txt

Installation

1. Install PacketEvents
2. Drop NullAntiCheat.jar into /plugins
3. Restart the server
4. Start in collect mode and gather labeled fights
5. Train model.json with the included script
6. Place model at plugins/NullAntiCheat/model/model.json
7. Switch to detect (or hybrid) and open /nac menu

Configuration highlights

mode: collect | hybrid | detect
dataset.critical-only: true
model.file: model/model.json
suspicion.recent-hits / thresholds
gui title, rows, filler

Messages are fully editable in messages.yml.



What's included

• NullAntiCheat plugin JAR
• config.yml + messages.yml
• Python trainer (train_model.py + requirements.txt)
• Setup / usage documentation

What's NOT included / not claimed

• Not a full replacement for GrimAC / Vulcan / Matrix
• No movement prediction / scaffold / fly suite (yet)
• No automatic bans out of the box
• Detection quality depends on your dataset quality
• A random public model will not magically fit every server

Recommended workflow

• Same arena, similar kits
• Record both legit and obvious kill aura under similar conditions
• Vary distance, strafing, jumping, combo pressure, ping
• Flush dataset, retrain, review false positives in /nac menu
• Tune thresholds only after real staff review

Support
After purchase you get:

• Installation help
• Trainer usage help
• Bugfix priority for the sold build

FAQ
Q: Does it ban players automatically?
A: No. It scores combat and shows suspects for staff. You decide the punishment flow.

Q: Do I need a GPU?
A: No. CatBoost training runs on CPU.

Q: Can I use it with another anticheat?
A: Yes. It is designed as a combat ML layer alongside existing AC plugins.

Q: Will one universal model work everywhere?
A: Best results come from training on your own labeled fights.

Q: Is this only for kill aura?
A: Yes — Stage 1 focus is behavioral kill aura / aim combat scoring.

NullAntiCheat — data first, decisions second.​