FalconEye : Real-time detection software for Windows process injections

FalconEye is a windows endpoint detection software for real-time process injections. It is a kernel-mode driver that aims to catch process injections as they are happening (real-time). Since FalconEye runs in kernel mode, it provides a stronger and reliable defense against process injection techniques that try to evade various user-mode hooks.

You can check our presentation at 2021 Blackhat ASIA Arsenal and slides.

Project Overview

Detection Coverage

The table below shows the implementation status and the detection logic for the various process injection techniques. WPM stands for WriteProcessMemory. To test the detection, one can refer to the references section.

TechniqueStatusDetectionPOC Used
AtombombingHook QueueUserAPC and look for GlobalGetAtom family of functionsPinjectra
Instrumentation callback injectionDetect if a new thread is created from floating code
Reflective DLL InjectionDetect if a new thread is created from floating code and if PE header is being written into victimMInjector
PROPagateHook SetProp to get the address of the property being written and corelate with the previous WPM calls to get the address of floating codePinjectra
Process HollowingDetected using PE header written into target process memoryMInjector
CreateRemoteThread with LoadLibraryNew thread with start address pointing to LoadLibrary. MInjector version also writes DLL path using WPM which is also detectedMInjector, Pinjectra
CreateRemoteThread with MapViewOfFileDetect if a new thread is created from floating codePinjectra
Suspend-Inject-ResumeDetect if a new thread is created from floating code(MInjector). DLL Path being written via WPM (MInjector). Detect if context set on a previously suspended thread (Pinjectra)MInjector, Pinjectra
QueueUserAPCDLL path being written via WPMMInjector
QueueUserAPC with memset (Stackbombing)Hook QueueUserAPC and look for memsetPinjectra
SetWindowLong (Extra window memory injection)Hook SetWindowLong to get the address of the function pointer being written and corelate with the previous WPM calls to get the address of floating codePinjectra
Unmap + OverwriteAlert if attacker process is unmapping ntdll from the victimPinjectra
Kernel Ctrl TableDetect if WPM is overwriting KernelCallbackTable field in the PEB of the victim
USERDATACheck if WPM target address is in conhost.exe range. If so check if any relevant function pointers from conhost match previously stored WPM address
Ctrl-injectDetect if the attacker does WPM in victim’s KernelBase.dll rangePinjectra
ALPC CallbackExtract victim pid in NtConnectPort calls to ALPC port. For attacker-victim pid tuple check prior WPM calls and apply Floating code detectionPinjectra
WNF CallbackWPM followed by UpdateWNFStateData call
SetWindowsHookSave module paths registered in NtUserSetWindowsHookEx hook. Later when a module matching this path loads in a different process, generate alertMInjector
GhostWritingDetect if context is set (NtSetContextThread is called) on a previously suspended threadPinjectra
Service ControlWPM overwriting Service IDE of a process (service)
Shellcode injectionNew thread started from floating code. DLL path being written by WPMMInjector
Image MappingThread started from floating code. PE header being written by WPM. DLL path being written by WPMMInjector
Thread ReuseThread started from floating code. DLL path being written by WPMMInjector

Architecture Overview

  • The driver is an on-demand load driver
  • The initialization includes setting up callbacks and syscall hooks via libinfinityhook
  • The callbacks maintain a map of Pids built from cross process activity such as OpenProcess but it is not limited to OpenProcess
  • Subsequent callbacks and syscall hooks use this Pid map to reduce the noise in processing. As a part of noise reduction, syscall hooks filter out same process activity.
  • The detection logic is divided into subcategories namely – stateless (example: Atombombing), stateful (Unmap+Overwrite) and Floating code(Shellcode from multiple techniques)
  • For stateful detections, syscall hooks record an ActionHistory which is implemented as a circular buffer. e.g. It records all the NtWriteVirtualMemory calls where the caller process is different from the target process.
  • The detection logic has common anomaly detection functionality such as floating code detection and detection for shellcode triggers in remote processes. Both callbacks and syscall hooks invoke this common functionality for actual detection.

NOTE: Our focus has been detection and not creating a performant detection engine. We’ll continue on these efforts past the BlackHat presentation.


├── src
│ ├── FalconEye —————————# FalconEye user and kernel space
│ └── libinfinityhook ———————# Kernel hook implementation
├── 2021BHASIA_FalconEye.pdf

Getting Started


  • Windows 10 Build 1903/1909
  • Microsoft Visual Studio 2019 onwards
  • Virtualization Software such as VmWare, Hyper-V (Optional)



  • Open the solution with Visual Studio 2019
  • Select x64 as build platform
  • Build solution. This should generate FalconEye.sys binary under src\kernel\FalconEye\x64\Debug or src\kernel\FalconEye\x64\Release

Test Machine Setup

  • Install Windows 10 Build 1903/1909 in a VM
  • Configure VM for testing unsigned driver
  • Using bcdedit, disable integrity checks : BCDEDIT /set nointegritychecks ON
  • Run DbgView from sysinternals in the VM or start a debugging connection using WinDbg.


  • Copy FalconEye.sys to the Test Machine (Windows 10 VM)
  • Load FalconEye.sys as ‘On Demand’ load driver using OSR Loader or similar tools
  • Run injection test tools such as pinjectra, minjector or other samples
  • Monitor debug logs either via WinDbg or DbgView