Game Hacking Techniques

- World-to-Screen transformation
- Player/entity rendering
- Box ESP, skeleton ESP
- Item highlighting
- Radar/minimap hacks

- Aimbot algorithms (memory-based and AI visual)
- Triggerbot (auto-fire on crosshair detection)
- No recoil/no spread
- Bullet prediction and lead calculation
- Silent aim (server-side angle manipulation)
- AI visual aimbot (YOLO-based, no memory access required)

Architecture overview:
Screen-capture paradigm — uses frame capture, object detection, and input
injection. Some implementations can avoid process attachment, a cheat driver,
and direct game-memory reads; that does not make the full pipeline artifact-free.

Typical setup:
┌─────────────────┐     screen capture      ┌──────────────────┐
│  Gaming PC      │ ───────────────────────▶ │  AI Pipeline     │
│  Game + OBS     │                          │  (same PC, or    │
│                 │ ◀─────────────────────── │   second PC)     │
└─────────────────┘     hardware input       │  YOLO model      │
                        (KMBox / Logitech)   │  TensorRT/CUDA   │
                                             └──────────────────┘

Dual-machine variant (separate processing location):
- Machine A (game): only runs game + OBS, sends frames via NDI/capture card
- Machine B (cheat): runs AI model, sends mouse commands via USB/network
  to hardware input device on Machine A
- Game machine need not run the model or decision logic, though capture,
  transport, and input-device artifacts can remain

Single-machine variant:
- OBS + AI model run on the same PC
- AI implemented as OBS filter plugin (looks like "OBS is running")
- Mouse output via hardware device or driver-level injection

Pipeline stages:

1. Frame Source Evidence:
   - Identify the actual source/backend and retained frame stage; a source label
     does not establish hook use, display coverage or a fixed frame rate
   - Treat capture plugins and external capture devices as separate provenance
     questions, and compare against legitimate recording configurations
   - Evaluate capture semantics using the graphics-api skill before interpreting
     a missing image or a process/module observation

2. AI Object Detection:
   - Model: YOLOv5 / YOLOv8 / YOLOv10 / YOLO11 (lightweight variants)
   - Training: fine-tuned on game-specific screenshots
     (enemy bodies, heads, torsos as labeled bounding boxes)
   - Input: cropped region around crosshair (320x320 or 640x640)
     to reduce inference cost
   - Output: bounding boxes with class (head/body/enemy) + confidence score
   - Acceleration: TensorRT (NVIDIA), CUDA, DirectML, OpenVINO
   - Set and measure the latency budget on the target capture path, model,
     hardware, frame rate, and input transport

3. Coordinate Transform and Aiming Logic:
   - Convert pixel coordinates to mouse movement delta:
     delta_x = (target_x - screen_center_x) * sensitivity
     delta_y = (target_y - screen_center_y) * sensitivity
   - Target selection: closest to crosshair, highest confidence,
     head priority, or combined scoring
   - FOV (Field of View) lock: only engage targets within
     configurable pixel radius from crosshair center

4. Attempts to mask automated trajectories:
   - Gradual movement with an acceleration curve instead of an instant snap
   - Synthetic jitter
   - Bézier curve or cubic interpolation for path
   - End-point correction (overshoot then settle)
   - Configurable engagement probability
   - Slight intentional offset (not pixel-perfect center-mass)
   - Variable reaction delay
   These transformations do not establish human equivalence; repeated
   parametric behavior can itself become a feature.

5. Mouse Movement Execution:
   - Hardware input devices (see Input Simulation section below)
   - Movement commands sent as physical HID reports
   - The host receives protocol-conformant HID input rather than a user-mode
     injection API call; device provenance and behavior may still be observable

Why OBS specifically:
- Legitimate streaming software, used by millions of streamers
- Blanket action against OBS-related processes would create substantial
  collateral impact; process presence alone is not attribution
- Game Capture provides fast, low-latency frame access
- Plugin system can host filters inside OBS, but loaded plugins, behavior, and
  surrounding telemetry may still be inspected
- Supports D3D11, D3D12, Vulkan, OpenGL capture paths

End-to-end workflow from raw game screenshots to deployed TensorRT model.

1. Data Collection:
   - Capture game screenshots during actual gameplay (OBS recording or replay)
   - Capture diverse scenarios: different maps, lighting, character skins,
     distances, poses, partial occlusion, smoke/flash effects
   - Determine dataset size from coverage and learning curves; image count alone
     does not guarantee robustness
   - Include negative samples (empty scenes, friendlies, environment objects)

2. Annotation / Labeling:
   - Tools: LabelImg (YOLO format), CVAT (collaborative), Roboflow (cloud),
     Label Studio, makesense.ai (browser-based)
   - YOLO format: one .txt per image, each line:
     <class_id> <center_x> <center_y> <width> <height>
     (all values normalized to 0-1 relative to image dimensions)
   - Class definitions (typical):
     0: enemy_body (full body bounding box)
     1: enemy_head (head-only bounding box, for headshot targeting)
     2: friendly (to avoid shooting teammates)
   - Label head separately from body for head-priority targeting
   - Quality control: consistent label boundaries, no missed instances

3. Data Augmentation:
   - Built-in Ultralytics augmentations (mosaic, mixup, copy-paste)
   - Game-specific augmentations:
     - Brightness/contrast variation (simulate different map lighting)
     - Random crop around crosshair area (match inference ROI)
     - Motion blur (simulate fast movement)
     - Noise injection (simulate compression artifacts)
   - Avoid augmentations that distort aspect ratio
     (characters would look unnatural, hurting accuracy)

4. Training:
   - Framework: Ultralytics YOLOv8/v10/v11/YOLO11
   - Base model: yolov8n.pt or yolov8s.pt (nano/small for speed)
     or yolo11n.pt for latest architecture
   - Training command:
     yolo detect train data=game_dataset.yaml model=yolov8n.pt
       epochs=100 imgsz=640 batch=16 device=0
   - dataset.yaml structure:
     path: /path/to/dataset
     train: images/train
     val: images/val
     names: {0: enemy_body, 1: enemy_head, 2: friendly}
   - Key hyperparameters to tune include input size, learning rate, confidence
     threshold, NMS IoU threshold, batch size, and augmentation policy
   - Measure training and inference cost on the exact model, software stack,
     precision, and target hardware

5. Validation and Testing:
   - Evaluate mAP@0.5 and mAP@0.5:0.95 on validation set
   - Choose operating thresholds from precision/recall and downstream error
     costs; no single mAP cutoff establishes reliable deployment
   - Test inference speed on target hardware and evaluate held-out maps, skins,
     patches, capture paths, and hard negatives

6. Export to TensorRT (deployment):
   - Step 1: Export to ONNX
     yolo export model=best.pt format=onnx simplify=True opset=17
   - Step 2: Convert ONNX to TensorRT engine
     yolo export model=best.pt format=engine half=True device=0
     (half=True enables FP16 precision)
   - Or use trtexec directly:
     trtexec --onnx=best.onnx --saveEngine=best.engine
       --fp16 --workspace=4096
   - Benchmark FP16 against FP32 on the exported model; latency, throughput, and
     accuracy changes are hardware- and graph-specific
   - INT8 can improve throughput but requires representative calibration data
     and accuracy validation

7. Runtime Integration:
   - Load TensorRT engine in C++/Python inference loop
   - Input: preprocessed frame (resize, normalize, HWC→CHW, float32/16)
   - Decode the exporter/version-specific output tensor; shapes and NMS
     placement vary across model and runtime versions
   - Apply NMS (Non-Maximum Suppression) to deduplicate detections
   - Select target based on: closest to crosshair + highest confidence
   - Convert pixel coordinates to mouse delta

Alternative acceleration backends:
- DirectML (AMD GPUs, Windows native)
- OpenVINO (Intel GPUs/CPUs)
- ONNX Runtime with CUDA EP (cross-platform)
- CoreML (macOS, less common for game cheats)

- Speed hacks
- Fly hacks
- No clip
- Teleportation
- Bunny hop automation

- Wallhacks
- Skin changers
- Unlock all
- Economy manipulation

- OpenProcess + ReadProcessMemory
- NtReadVirtualMemory
- Memory-mapped files
- Shared memory sections

- Driver-based access
- Physical memory access
- MDL-based copying
- KeStackAttachProcess

- DMA (Direct Memory Access)
- EFI runtime services
- Hypervisor-based access
- Hardware-based (FPGA)

- Disk serial: IOCTL_STORAGE_QUERY_PROPERTY, SMART data
- NIC MAC address: NDIS OID_802_3_PERMANENT_ADDRESS
- SMBIOS: motherboard serial, system UUID, BIOS vendor
- GPU serial: registry-based or NVAPI/ADL queries
- Monitor EDID: display serial number
- Volume serial: NtQueryVolumeInformationFile
- TPM EK: Endorsement Key fingerprint

- Disk filter driver: intercept IOCTL and replace serial in response
- Registry value spoofing: modify cached hardware IDs
- SMBIOS table patching: modify raw SMBIOS memory region
- NIC driver hook: replace MAC in NDIS miniport response
- Full HWID spoofer: coordinated spoofing across all identifiers

- Replace return address on stack before API call
- Restore original after call returns
- Evades stack-walk-based detection (RtlWalkFrameChain)
- Techniques: JMP RBX gadget, synthetic frames, fiber-based

- Build fake but plausible call stack frames
- Match expected module return addresses (ntdll, kernel32)
- Evade NtQueryInformationThread stack inspection
- Tools: SpoofCallStack, Vulcan, CallStackSpoofer

- Anti-cheat walks thread stacks looking for non-module returns
- Stack unwinding via .pdata / UNWIND_INFO validation
- Spoofed stacks must pass RtlVirtualUnwind consistency checks

IOCTL-based:
- Standard DeviceIoControl with custom control codes
- Buffered I/O, Direct I/O, METHOD_NEITHER

Data pointer swaps (abusing legitimate syscalls):
- NtUserGetObjectInformation
- NtConvertBetweenAuxiliaryCounterAndPerformanceCounter
- NtUserRegisterRawInputDevices
- NtGdiGetCOPPCompatibleOPMInformation
- NtDxgkGetTrackedWorkloadStatistics
- NtUserGetPointerInfoList
- NtUserSetInformationThread
- NtDCompositionSetChildRootVisual
- Win32k syscall hooks

Shared memory:
- Named shared sections (ZwCreateSection + ZwMapViewOfSection)
- Physical memory mapping
- Shared event objects for signaling

Callback-based:
- Registry callbacks (CmRegisterCallbackEx)
- Minifilter communication ports (FltCreateCommunicationPort)
- Object callbacks with embedded data

Unconventional channels:
- Named pipes from kernel
- Window messages (NtUserPostMessage)
- ETW provider channels
- Socket from kernel (Winsock Kernel / WSK)
- File system filter callbacks
- Debugging APIs (DbgPrint interception)

Vector2 WorldToScreen(Vector3 worldPos, Matrix viewMatrix) {
    Vector4 clipCoords;
    clipCoords.x = worldPos.x * viewMatrix[0] + worldPos.y * viewMatrix[4] + 
                   worldPos.z * viewMatrix[8] + viewMatrix[12];
    clipCoords.y = worldPos.x * viewMatrix[1] + worldPos.y * viewMatrix[5] + 
                   worldPos.z * viewMatrix[9] + viewMatrix[13];
    clipCoords.w = worldPos.x * viewMatrix[3] + worldPos.y * viewMatrix[7] + 
                   worldPos.z * viewMatrix[11] + viewMatrix[15];
    
    if (clipCoords.w < 0.1f) return invalid;
    
    Vector2 NDC;
    NDC.x = clipCoords.x / clipCoords.w;
    NDC.y = clipCoords.y / clipCoords.w;
    
    Vector2 screen;
    screen.x = (screenWidth / 2) * (NDC.x + 1);
    screen.y = (screenHeight / 2) * (1 - NDC.y);
    
    return screen;
}

KMBox Net (network variant) — UDP-based protocol:

Packet header (16 bytes, Little-Endian):
Offset  Field      Size   Description
0x00    MAC        4 B    Device UUID (unique per device, used for auth)
0x04    RAND       4 B    Random value or parameter
0x08    INDEXPTS   4 B    Incrementing sequence number (replay protection)
0x0C    CMD        4 B    Command code

Key command codes:
The values below are firmware/API-version examples; verify them against the
exact device implementation before analysis.
Code          Command          Description
0xAF3C2828    connect          Establish connection with device
0xAEDE7345    mouse_move       Direct mouse movement (dx, dy)
0xAEDE7346    mouse_automove   Human-like movement with interpolation
0xA238455A    mouse_beizer     Bézier curve mouse movement
0x9823AE8D    mouse_left       Left button press/release
0x238D8212    mouse_right      Right button press/release
0x97A3AE8D    mouse_middle     Middle button press/release
0xFFEEAD38    mouse_wheel      Scroll wheel
0x123C2C2F    keyboard_all     Keyboard key event

Mouse API functions:
- move(x, y):                 Direct relative movement, no interpolation
- move_auto(x, y, ms):        Interpolated movement over ms milliseconds
- move_beizer(x, y, ms,       Second-order Bézier curve with custom
    x1, y1, x2, y2):          control points for trajectory shaping

Encrypted variants (enc_*):   Same functions with packet-level encryption
                               to resist network packet analysis

Performance:
- Measure command rate, latency distribution, loss, buffering, and jitter on
  the exact firmware, transport, host, and network; fixed figures do not
  transfer across setups

KMBox B / B Pro (serial variant):
- USB CDC serial communication (COM port)
- Baud rate is firmware/configuration-specific (115200 is one common setting)
- Simpler protocol: ASCII or binary command frames
- Benchmark round-trip timing on the deployed serial stack

Physical keyboard/mouse monitoring:
- monitor() function reads real user input from the device
- Enables "pass-through + inject" mode:
  real user input flows through normally,
  AI-calculated deltas are added on top

Arduino / Teensy HID protocol:
- Custom serial command format (typically simple ASCII):
  "M,dx,dy\n"      — mouse move
  "C,button\n"      — click (1=left, 2=right, 3=middle)
  "K,keycode\n"     — keypress
- USB HID report generated by ATmega32U4 (Leonardo)
  or ARM-based Teensy (3.2, 4.0, 4.1)
- HID report descriptor mimics standard mouse:
  buttons (3 bits) + X delta (8-16 bits) + Y delta (8-16 bits)
- No custom driver needed — OS uses generic HID driver

Logitech driver API (exploitable versions):
- G HUB versions prior to certain patches expose internal functions
- Key DLLs: LGS (lcore.dll), G HUB (ghub_mouse.dll or internal APIs)
- ghub_mouse_move(dx, dy) or equivalent internal symbol
- Accessed via DLL injection into GHUB process
  or LoadLibrary + GetProcAddress
- Movement appears as Logitech device input in the HID stack
- Patched in newer G HUB versions; specific version numbers
  circulate in cheat communities

1. Pattern scan for signatures
2. Read game memory externally
3. Process data in separate process
4. Render overlay or use input simulation

1. Inject into game process
2. Hook rendering functions
3. Access game objects directly
4. Render through game's graphics context

description/{owner}/{repo}/description_en.txt
https://raw.githubusercontent.com/gmh5225/awesome-game-security/refs/heads/main/description/{owner}/{repo}/description_en.txt

archive/{owner}/{repo}.txt
https://raw.githubusercontent.com/gmh5225/awesome-game-security/refs/heads/main/archive/{owner}/{repo}.txt