Skip to content

Division-36/Planck-99_PublicBenchmarks

Repository files navigation

V2 benchmarks in v2 directory

Why This Matters — Planck-99

  • Planck-99 is a kernel-level malware detection system that enables real-time threat detection on devices that previously could not run security software.
  • It reduces detection latency from milliseconds to nanoseconds and operates within ~37KB, making it viable for IoT, embedded systems, and low-resource environments.
  • This allows organizations to deploy behavioral security directly on-device — eliminating reliance on cloud analysis and reducing response time from seconds to instant.

The Problem

Modern malware detection does not operate where it matters most.

Existing solutions are:

  • Too heavy for embedded / IoT systems
  • Too slow for real-time detection
  • Dependent on long traces or post-execution analysis

As a result, vast numbers of devices operate without effective protection at the behavioral level.


The Solution

Planck-99 is a syscall-based malware classifier designed for real-time, embedded environments.

It operates with:

  • ~30–70 nanoseconds inference latency
  • ~37 KB total system footprint
  • No dependency on trace length at inference time

The model runs directly at the system level and enables continuous behavioral monitoring.


What Makes It Different

1. Length-Invariant Detection

Traditional approaches degrade or slow down with longer traces.

Planck-99 uses normalized syscall frequency ratios, making detection:

  • Independent of trace length
  • Stable across vastly different execution scales

This enables generalization up to 51× beyond training conditions.


2. Real-Time Kernel-Level Execution

The system is implemented as a lightweight C kernel component:

  • Deterministic latency (nanoseconds)
  • No runtime overhead accumulation
  • Suitable for continuous monitoring loops

This is not post-analysis — it is inline detection.


3. Embedded-First Design

Planck-99 is designed for environments where ML typically fails:

  • IoT devices
  • Low-memory systems
  • Edge deployments without accelerators

Total footprint fits within tens of kilobytes.


4. Explicit Failure Modeling

The system identifies and accounts for its limitations:

  • Short traces (<500 syscalls) produce unstable signals
  • A gating mechanism prevents unreliable classification
  • Sliding window ensures continuous accuracy over time

This makes the system predictable and controllable in production


Why It Matters

Planck-99 shifts malware detection from:

  • Heavy → lightweight
  • Delayed → real-time
  • Server-dependent → device-native

This enables a new deployment model:

Behavioral security at the lowest levels of computation.

Devices that previously could not run security models can now:

  • Detect threats locally
  • Respond instantly
  • Operate without external dependencies

The Outcome

  • Orders-of-magnitude faster inference (ns vs ms)
  • Minimal resource usage (KB vs MB/GB)
  • Strong generalization across datasets and time

Planck-99 redefines where malware detection can exist — moving it from the cloud to the device itself. It is a change in where and how malware detection can exist.


Status

Planck-99 is currently under active testing and benchmarking across multiple datasets, including unseen distributions.

All results, benchmarks, and system details are publicly available in this repository.


Intellectual Property Notice:
Planck-99 is a proprietary technology developed by Ziad Salah under the Division-36 umbrella.
All rights reserved. Commercial licensing and partnership inquiries: zs.01117875692@gmail.com

Bottom Line

If malware detection is to scale across billions of devices,
it must become:

  • Lightweight
  • Deterministic
  • Always-on

Planck-99 is a step in that direction.

Releases

No releases published

Packages

 
 
 

Contributors

Languages