README.md

# TFLite MEAN Operator Stack Buffer Overflow via ResolveAxis()

## Vulnerability

The TFLite MEAN operator's `ResolveAxis()` function at `reduce.cc:425` copies
axis tensor elements from the model file into a fixed-size `MeanParams.axis[4]`
array (4 elements, 8 bytes) on the stack without checking bounds. A crafted
model with more than 4 axis elements causes a stack buffer overflow with
attacker-controlled write size.

**CWE-121: Stack-based Buffer Overflow**

## Impact

- **ASAN build:** Crashes at the first out-of-bounds write (detected by ASAN shadow memory)
- **Production build (Python `tensorflow` 2.20.0):** Crashes with SIGSEGV or SIGFPE depending on overflow depth
- The overflow size is fully controlled by the attacker (axis tensor element count in the model file)
- With 200 axis elements: 392 bytes of int16 values overwrite adjacent stack variables, saved frame pointer, and return address area

## Reproduction

```bash
pip install tensorflow-cpu  # or tensorflow
python3 poc_generator.py 200
python3 -c "
import tensorflow as tf
interp = tf.lite.Interpreter(model_path='/tmp/mean_rce_axis200.tflite')
interp.allocate_tensors()
interp.invoke()  # crashes with SIGSEGV or SIGFPE
"
```

Or run `./reproduce.sh`.

## Files

- `poc_generator.py` — Generates crafted TFLite model with MEAN op and large axis tensor
- `poc_axis200.tflite` — Pre-generated PoC model (200 axis elements, 1360 bytes)
- `reproduce.sh` — Self-contained reproduction script
- `asan_output.txt` — Full ASAN crash trace

## Root Cause

In `tensorflow/lite/kernels/reduce.cc:420-427`:

```cpp
void ResolveAxis(const int* axis_data, int axis_count,
                 tflite::MeanParams* op_params) {
  int i = 0;
  for (; i < axis_count; ++i) {
    op_params->axis[i] = static_cast<int16>(axis_data[i]);  // No bounds check!
  }
  // ...
}
```

`MeanParams.axis` is declared in `types.h` as `int16_t axis[4]` (only 4 elements).
`axis_count` comes from `NumElements(op_context.axis)` which reads the axis tensor
shape from the model file — fully attacker-controlled with no validation against
the array size.

## Trigger Path

```
Model.Invoke()
  → Subgraph::InvokeImpl()
    → reduce::EvalMean<kGenericOptimized>()  [reduce.cc:551]
      → case kTfLiteUInt8:  [reduce.cc:611]
        → ResolveAxis(axis_data, num_axis, &op_params)  [reduce.cc:613]
          → op_params.axis[i] = ... for i=0..199  ← OVERFLOW at i=4
```

Requires UINT8 input tensor to reach the vulnerable code path.