VYPR
← All advisories
High severityNVD Advisory· Published Aug 12, 2021· Updated Aug 4, 2024

Incomplete validation in MKL requantization in TensorFlow

CVE-2021-37665

Description

TensorFlow is an end-to-end open source platform for machine learning. In affected versions due to incomplete validation in MKL implementation of requantization, an attacker can trigger undefined behavior via binding a reference to a null pointer or can access data outside the bounds of heap allocated arrays. The implementation does not validate the dimensions of the input tensor. A similar issue occurs in MklRequantizePerChannelOp. The implementation does not perform full validation for all the input arguments. We have patched the issue in GitHub commit 9e62869465573cb2d9b5053f1fa02a81fce21d69 and in the Github commit 203214568f5bc237603dbab6e1fd389f1572f5c9. The fix will be included in TensorFlow 2.6.0. We will also cherrypick this commit on TensorFlow 2.5.1, TensorFlow 2.4.3, and TensorFlow 2.3.4, as these are also affected and still in supported range.

Affected packages

Versions sourced from the GitHub Security Advisory.

PackageAffected versionsPatched versions
tensorflowPyPI
< 2.3.42.3.4
tensorflowPyPI
>= 2.4.0, < 2.4.32.4.3
tensorflowPyPI
>= 2.5.0, < 2.5.12.5.1
tensorflow-cpuPyPI
< 2.3.42.3.4
tensorflow-cpuPyPI
>= 2.4.0, < 2.4.32.4.3
tensorflow-cpuPyPI
>= 2.5.0, < 2.5.12.5.1
tensorflow-gpuPyPI
< 2.3.42.3.4
tensorflow-gpuPyPI
>= 2.4.0, < 2.4.32.4.3
tensorflow-gpuPyPI
>= 2.5.0, < 2.5.12.5.1

Affected products

1

Patches

2
203214568f5b

Reorganize and add more validation to MKL requantization

https://github.com/tensorflow/tensorflowMihai MaruseacJul 30, 2021via ghsa
commit
1 file changed · +25 −15
  • tensorflow/core/kernels/mkl/mkl_requantize_per_channel_op.cc+25 −15 modified
    @@ -49,35 +49,45 @@ class MklRequantizePerChannelOp : public OpKernel {
   void Compute(OpKernelContext* ctx) override {
     try {
       const Tensor& input = ctx->input(kInputTensorIndex);
+      OP_REQUIRES(
+          ctx, input.dims() == 4,
+          errors::InvalidArgument("Current RequantizePerChannel operator"
+                                  "supports 4D tensors only."));
+
       const Tensor& input_min_vec = ctx->input(kInputMinVecIndex);
+      size_t depth = input_min_vec.NumElements();
       float* input_min_vec_data = (float*)const_cast<void*>(
           static_cast<const void*>(input_min_vec.flat<float>().data()));
+
       const Tensor& input_max_vec = ctx->input(kInputMaxVecIndex);
+      OP_REQUIRES(
+          ctx, input_max_vec.NumElements() == depth,
+          errors::InvalidArgument("input_max has incorrect size, expected ",
+                                  depth, " was ", input_max_vec.NumElements()));
       float* input_max_vec_data = (float*)const_cast<void*>(
           static_cast<const void*>(input_max_vec.flat<float>().data()));
 
       const Tensor& input_requested_min = ctx->input(this->kRequestMinIndex);
+      OP_REQUIRES(
+          ctx, input_requested_min.NumElements() == 1,
+          errors::InvalidArgument("requested_output_min must be a scalar"));
       const float input_requested_min_float =
           input_requested_min.flat<float>()(0);
+
       const Tensor& input_requested_max = ctx->input(this->kRequestMaxIndex);
+      OP_REQUIRES(
+          ctx, input_requested_min.NumElements() == 1,
+          errors::InvalidArgument("requested_output_max must be a scalar"));
       const float input_requested_max_float =
           input_requested_max.flat<float>()(0);
 
-      size_t depth = input_min_vec.NumElements();
-      OP_REQUIRES(
-          ctx, input.dims() == 4,
-          errors::InvalidArgument("Current RequantizePerChannel operator"
-                                  "supports 4D tensors only."));
-      OP_REQUIRES(
-          ctx, input_min_vec.dim_size(0) == depth,
-          errors::InvalidArgument("input_min has incorrect size, expected ",
-                                  depth, " was ", input_min_vec.dim_size(0)));
-      OP_REQUIRES(
-          ctx, input_max_vec.dim_size(0) == depth,
-          errors::InvalidArgument("input_max has incorrect size, expected ",
-                                  depth, " was ", input_max_vec.dim_size(0)));
-
-      if (out_type_ == DT_QINT8) DCHECK(input_requested_min_float < 0.0f);
+      if (out_type_ == DT_QINT8) {
+        OP_REQUIRES(ctx, input_requested_min_float < 0.0f,
+                    errors::InvalidArgument(
+                        "If out_type is QINT8, requested_output_max must be "
+                        "non negative, got ",
+                        input_requested_min_float));
+      }
 
       const float factor = (out_type_ == DT_QINT8) ? 127.0f : 255.0f;
       const float requested_min_max =
9e6286946557

Add more validation to `RequantizationRangePerChannel`.

https://github.com/tensorflow/tensorflowMihai MaruseacJul 29, 2021via ghsa
commit
1 file changed · +14 −0
  • tensorflow/core/kernels/mkl/mkl_requantization_range_per_channel_op.cc+14 −0 modified
    @@ -57,6 +57,20 @@ class MklRequantizationRangePerChannelOp : public OpKernel {
         ctx, input_max.dim_size(0) == depth,
         errors::InvalidArgument("input_max has incorrect size, expected ",
                                 depth, " was ", input_max.dim_size(0)));
+    OP_REQUIRES(
+        ctx, input_min.NumElements() == depth,
+        errors::InvalidArgument("input_min must have the same number of "
+                                "elements as input_max, got ",
+                                input_min.NumElements(), " and ", depth));
+    OP_REQUIRES(ctx, input.NumElements() > 0,
+                errors::InvalidArgument("input must not be empty"));
+    OP_REQUIRES(ctx, input.dims() == 4,
+                errors::InvalidArgument("input must be in NHWC format"));
+    OP_REQUIRES(
+        ctx, input.dim_size(3) == depth,
+        errors::InvalidArgument(
+            "input must have same number of channels as length of input_min: ",
+            input.dim_size(3), " vs ", depth));
 
     const float* input_min_data = input_min.flat<float>().data();
     const float* input_max_data = input_max.flat<float>().data();

Vulnerability mechanics

Generated by null/stub on May 9, 2026. Inputs: CWE entries + fix-commit diffs from this CVE's patches. Citations validated against bundle.

References

8

News mentions

0

No linked articles in our index yet.