renaissancefieldlite/tris-keras-native-architectural-backdoor-poc
014
Tris Keras Native Architectural Backdoor PoC
This repository is a minimal responsible-security proof of concept for a Keras Native .keras model-integrity issue.
The model uses only built-in Keras graph operations. It loads with safe_mode=True, behaves normally for benign input, and preserves a dormant trigger branch that overrides the output for a specific input threshold.
Behavior
- Normal input:
12.0 -> 12.0 - Trigger input:
67.0 -> 999.0 - Native save: passed
- Safe-mode load: passed
Reproduce
KERAS_BACKEND=torch python build_keras_native_arch_backdoor_poc.pyBoundary
This is not a parser memory-corruption bug and does not execute operating system commands. It is a model-integrity / architectural-backdoor PoC showing that hidden trigger behavior can be preserved inside a valid Keras Native model loaded with safe mode.
Reported through Huntr by Renaissance Field Lite.
