Headline
CVE-2022-35972: Fix quantize ops input validation issues. · tensorflow/tensorflow@785d67a
TensorFlow is an open source platform for machine learning. If QuantizedBiasAdd is given min_input, max_input, min_bias, max_bias tensors of a nonzero rank, it results in a segfault that can be used to trigger a denial of service attack. We have patched the issue in GitHub commit 785d67a78a1d533759fcd2f5e8d6ef778de849e0. The fix will be included in TensorFlow 2.10.0. We will also cherrypick this commit on TensorFlow 2.9.1, TensorFlow 2.8.1, and TensorFlow 2.7.2, as these are also affected and still in supported range. There are no known workarounds for this issue.
@@ -53,10 +53,10 @@ TEST_F(RequantizeTest, HandCraftedRequantize) { // Requantize to -1 to 1. AddInputFromArray<qint32>(TensorShape({value_count}), {-(1 << 23), 0, (1 << 23)}); AddInputFromArray<float>(TensorShape({1}), {-256.0f}); AddInputFromArray<float>(TensorShape({1}), {256.0f}); AddInputFromArray<float>(TensorShape({1}), {-1.0f}); AddInputFromArray<float>(TensorShape({1}), {1.0f}); AddInputFromArray<float>(TensorShape({}), {-256.0f}); AddInputFromArray<float>(TensorShape({}), {256.0f}); AddInputFromArray<float>(TensorShape({}), {-1.0f}); AddInputFromArray<float>(TensorShape({}), {1.0f}); TF_ASSERT_OK(RunOpKernel()); Tensor expected(allocator(), DT_QUINT8, TensorShape({value_count})); test::FillValues<quint8>(&expected, {0, 128, 255}); @@ -71,10 +71,10 @@ TEST_F(RequantizeTest, InvalidOutputMin) {
AddInputFromArray<qint32>(TensorShape({value_count}), {-(1 << 23), 0, (1 << 23)}); AddInputFromArray<float>(TensorShape({1}), {-256.0f}); AddInputFromArray<float>(TensorShape({1}), {256.0f}); AddInputFromArray<float>(TensorShape({1}), {0.01f}); AddInputFromArray<float>(TensorShape({1}), {1.0f}); AddInputFromArray<float>(TensorShape({}), {-256.0f}); AddInputFromArray<float>(TensorShape({}), {256.0f}); AddInputFromArray<float>(TensorShape({}), {0.01f}); AddInputFromArray<float>(TensorShape({}), {1.0f}); EXPECT_EQ("requested_output_min must be <= 0, but got 0.01", RunOpKernel().error_message()); } @@ -85,10 +85,10 @@ TEST_F(RequantizeTest, InvalidOutputMax) {
AddInputFromArray<qint32>(TensorShape({value_count}), {-(1 << 23), 0, (1 << 23)}); AddInputFromArray<float>(TensorShape({1}), {-256.0f}); AddInputFromArray<float>(TensorShape({1}), {256.0f}); AddInputFromArray<float>(TensorShape({1}), {-10.0f}); AddInputFromArray<float>(TensorShape({1}), {-11.0f}); AddInputFromArray<float>(TensorShape({}), {-256.0f}); AddInputFromArray<float>(TensorShape({}), {256.0f}); AddInputFromArray<float>(TensorShape({}), {-10.0f}); AddInputFromArray<float>(TensorShape({}), {-11.0f}); EXPECT_EQ( "requested_output_max must be >= requested_output_min, but got -11 and " "-10",
Related news
### Impact If `QuantizedBiasAdd` is given `min_input`, `max_input`, `min_bias`, `max_bias` tensors of a nonzero rank, it results in a segfault that can be used to trigger a denial of service attack. ```python import tensorflow as tf out_type = tf.qint32 input = tf.constant([85,170,255], shape=[3], dtype=tf.quint8) bias = tf.constant(43, shape=[2,3], dtype=tf.quint8) min_input = tf.constant([], shape=[0], dtype=tf.float32) max_input = tf.constant(0, shape=[1], dtype=tf.float32) min_bias = tf.constant(0, shape=[1], dtype=tf.float32) max_bias = tf.constant(0, shape=[1], dtype=tf.float32) tf.raw_ops.QuantizedBiasAdd(input=input, bias=bias, min_input=min_input, max_input=max_input, min_bias=min_bias, max_bias=max_bias, out_type=out_type) ``` ### Patches We have patched the issue in GitHub commit [785d67a78a1d533759fcd2f5e8d6ef778de849e0](https://github.com/tensorflow/tensorflow/commit/785d67a78a1d533759fcd2f5e8d6ef778de849e0). The fix will be included in TensorFlow 2.10.0. We will also c...