HegzaPlacx/armnn-deser-poc

Arm NN .armnn deserializer — memory-safety PoC model files

Malicious Arm NN Flatbuffers (.armnn) model files that corrupt memory when loaded via armnnDeserializer::IDeserializer::CreateNetworkFromBinary(). Tested on ARM-software/armnn HEAD f8beb5a9 (src/armnnDeserializer/Deserializer.cpp).

File	Result	Sink
armnn_write_oob.armnn	stack out-of-bounds WRITE (CWE-787)	ToTensorInfo, Deserializer.cpp:756
armnn_read_oob.armnn	out-of-bounds READ (CWE-125, same root cause)	GetInputs/GetBaseLayer, Deserializer.cpp:839/286
armnn_valid.armnn	loads cleanly (negative control)	—

Root cause

CreateNetworkFromBinary runs the FlatBuffers structural verifier (VerifyGraphBuffer), which does not range-check scalar values. Attacker-controlled fields are then used without validation:

• ToTensorInfo writes dimensionsSpecificity[i] into a fixed bool[armnn::MaxNumOfTensorDimensions] (= bool[5]); i is bounded only by the attacker dimensionSpecificity vector length → stack overflow.
• GetInputs passes Connection.sourceLayerIndex through CHECKED_NON_NEGATIVE (lower bound only) into layers()->Get(index) → out-of-bounds.

In release builds (-DNDEBUG) the FlatBuffers debug assert is compiled out, so these are raw OOB accesses.

Reproduce

Self-contained harness (uses the project ArmnnSchema.fbs generated header + FlatBuffers headers):

c++ -std=c++17 -DNDEBUG -fsanitize=address -g -I. -Iflatbuffers_include harness_dimspec.cpp -o h_write
./h_write armnn_write_oob.armnn      # AddressSanitizer: stack-buffer-overflow WRITE in ToTensorInfo
./h_write armnn_valid.armnn          # clean

c++ -std=c++17 -DNDEBUG -fsanitize=address -g -I. -Iflatbuffers_include harness_read.cpp -o h_read
./h_read armnn_read_oob.armnn        # AddressSanitizer: heap-buffer-overflow READ in GetBaseLayer

Also reproducible via ExecuteNetwork built with -fsanitize=address: ExecuteNetwork -m armnn_write_oob.armnn -f armnn-binary.

Suggested fix

• ToTensorInfo: reject dimensionSpecificity->size() > MaxNumOfTensorDimensions.
• GetInputs/GetBaseLayer: add inputId < graph->layers()->size().

Crash-only proof of concept.