hacnho/coreml-traininginput-fanout-dos-poc

Core ML trainingInput materialization DoS PoC

This repository contains a benign security research PoC for a .mlmodel artifact that drives large protobuf materialization during Core ML spec loading.

Files:

• control_same_size.mlmodel
• malicious_traininginput_1000000.mlmodel
• reproduce.py

Observed behavior:

• control artifact:
  • parses successfully with zero trainingInput entries
• malicious artifact:
  • same size as control
  • parses successfully with 1,000,000 trainingInput entries
  • load_spec() peak RSS delta is about 119904 kB
  • MLModel(skip_model_load=True) peak RSS delta is about 228988 kB

Public files:

• https://huggingface.co/hacnho/coreml-traininginput-fanout-dos-poc/resolve/main/control_same_size.mlmodel
• https://huggingface.co/hacnho/coreml-traininginput-fanout-dos-poc/resolve/main/malicious_traininginput_1000000.mlmodel
• https://huggingface.co/hacnho/coreml-traininginput-fanout-dos-poc/resolve/main/reproduce.py

Reproduction:

python3 build_poc.py
python3 reproduce.py