You can deploy this model and start earning money today!

ericblackgachara/
armnn-heap-overflow-poc

Name: armnn-heap-overflow-poc
Availability: InStock

Views

No views yet

CWE-122: Heap Buffer Overflow in armnn TFLiteParser (INT64 Type Confusion)

Field	Value
Target	ARM-software/armnn
Version	v26.01 (HEAD f8beb5a)
CWE	CWE-122 Heap-Based Buffer Overflow
CVSS	8.8 HIGH (AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H)
Platform	huntr.com
Status	SUBMITTED

Root Cause (one line)

vector<unsigned int> (4 bytes/elem) allocated by GetNumElements(), then memcpy(GetNumBytes()) — diverge when attacker sets auxiliary tensor type = INT64 (8 bytes/elem), producing heap overflow of GetNumElements() * 4 bytes.

Trigger Path

ITfLiteParser::CreateNetworkFromBinaryFile("malicious.tflite")
  -> ParseTranspose / ParseBatchToSpaceNd / ParseSlice / ParseStridedSlice / ParsePad
      -> ValidateBuffer passes (checks file buffer size, not element type)
      -> vector<unsigned int>(N) allocated (16 bytes for N=4)
      -> memcpy(dest, src, GetNumBytes()) copies N*8=32 bytes
      -> HEAP BUFFER OVERFLOW (16 bytes past end)

Reproduce

bash

g++ -std=c++14 -fsanitize=address -g -O0 poc_tflite_type_confusion.cpp -o poc_tflite
./poc_tflite 1    # Transpose   -> ASAN: heap-buffer-overflow at TfLiteParser.cpp:1666
./poc_tflite 2    # BatchToSpaceNd -> ASAN: heap-buffer-overflow at TfLiteParser.cpp:1906
./poc_tflite 3    # StridedSlice -> ASAN: heap-buffer-overflow at TfLiteParser.cpp:2650
./poc_tflite 4    # Pad         -> ASAN: heap-buffer-overflow at TfLiteParser.cpp:3107

Expected output for each case:

ERROR: AddressSanitizer: heap-buffer-overflow
WRITE of size 32 at ... (0 bytes after 16-byte region)
SUMMARY: AddressSanitizer: heap-buffer-overflow ...

Affected Locations

Operation	File:Line	Vector type
ParseTranspose	TfLiteParser.cpp:1666	vector<unsigned int>
ParseBatchToSpaceNd blockShape	TfLiteParser.cpp:1906	vector<unsigned int>
ParseBatchToSpaceNd cropsVector	TfLiteParser.cpp:1909	vector<unsigned int>
ParseSlice begin	TfLiteParser.cpp:2206	vector<unsigned int>
ParseSlice signedSize	TfLiteParser.cpp:2220	vector<int>
ParseStridedSlice begin/end/stride	TfLiteParser.cpp:2650-2682	vector<int> x3
ParsePad	TfLiteParser.cpp:3107	vector<unsigned int>

ParseSpaceToBatchNd is NOT affected — has explicit DataType::Signed32 guard at line 2388.

Fix

Add the same guard already present in ParseSpaceToBatchNd to each affected operation:

cpp

if (auxTensorInfo.GetDataType() != armnn::DataType::Signed32)
    throw ParseException("Auxiliary tensor must be of type INT32.");

Files

File	Purpose
`poc_tflite_type_confusion.cpp`	ASAN harness — 4 cases, all confirmed
`submission.md`	Full markdown writeup (attachment)
`report.md`	Plain prose — huntr form paste target
`poc-evidence.html`	Dark terminal evidence page
`README.md`	This file

ericblackgachara/armnn-heap-overflow-poc