Harden OneHot operator input validation and output size computation

[GopalakrishnanN]

Harden OneHot operator input validation and output size computation

Summary

This PR tightens input validation and output-shape computation for the OneHot operator on the CPU and CUDA execution providers so that invalid or extreme inputs fail cleanly with an INVALID_ARGUMENT status instead of triggering overflow, silent truncation, or division by zero.

Changes

1. Overflow check in PrepareOutputShape using SafeInt

   • The output tensor size computation now uses a checked multiplication over the output dimensions, and the prefix_dim_size multiplication loop is rewritten with SafeInt<int64_t>. This prevents unbounded allocation attempts when a large depth value (or large indices shape) would overflow int64_t.

2. Guard against division by zero when prefix_dim_size is zero

   • suffix_dim_size is now computed as (prefix_dim_size > 0) ? (indices_shape.Size() / prefix_dim_size) : 0, avoiding an integer division by zero when an indices dimension before the axis is zero.

3. CUDA int32 range validation before fast_divmod

   • Before narrowing to int for fast_divmod (which requires int32 operands), the kernel now validates that suffix_dim_size and depth_val * suffix_dim_size fit in int32_t. Previously gsl::narrow_cast<int> could silently truncate, producing wrong divmod operands.

4. nullptr check on Output() in both CPU and CUDA Compute paths

   • If allocation of the output tensor fails (e.g., because the requested shape is too large for the allocator), both kernels now return a descriptive error instead of dereferencing a null pointer.

5. Unit tests in onnxruntime/test/providers/cpu/tensor/onehot_op_test.cc:

   • DepthTooLarge_OutputSizeOverflow — depth = INT64_MAX, indices = [2,3].
   • DepthTooLarge_OutputSizeOverflow_LargeIndices — depth = INT64_MAX / 500, indices = [1000].
   • NegativeDepth — negative depth is rejected.
   • DepthOne — minimum valid depth = 1 edge case.
   • ScalarIndicesRejected — rank-0 indices are rejected per the ONNX spec (indices rank ≥ 1).
   • DefaultAxis_Opset9 — opset 9 coverage for the default-axis path.
   • The three negative-path tests exclude kTensorrtExecutionProvider and kDmlExecutionProvider because those EPs fail with their own (different) error messages before our kernel validation runs.

Files touched

onnxruntime/core/providers/cpu/tensor/onehot.cc        | 28 +++++++-
onnxruntime/core/providers/cuda/tensor/onehot.cc       | 13 ++++
onnxruntime/test/providers/cpu/tensor/onehot_op_test.cc | 83 ++++++++++++++++++++++

Testing

Built onnxruntime_provider_test on Windows (Release) and ran OneHotOpTest.*:

[==========] 34 tests from 1 test suite ran.
[  PASSED  ] 34 tests.

Overflow tests produce the expected error, for example:

OneHot: output tensor size would overflow for the given indices shape and depth value (9223372036854775807).

[Copilot]

Pull request overview

Addresses a reported DoS risk in the OneHot operator by adding shape-size validation and extra allocation guarding to reduce the chance of oversized/overflowing output allocations during execution (CPU/CUDA).

Changes:

• Add an int64 element-count overflow check in PrepareOutputShape() for OneHot.
• Add output allocation null checks in CPU and CUDA OneHot compute paths.
• Add new unit tests intended to verify overflow rejection behavior.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
onnxruntime/core/providers/cpu/tensor/onehot.cc	Adds output element-count overflow validation and an output allocation null guard.
onnxruntime/core/providers/cuda/tensor/onehot.cc	Adds an output allocation null guard in the CUDA compute path.
onnxruntime/test/providers/cpu/tensor/onehot_op_test.cc	Adds two new failure-mode tests for extremely large depth values.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tianleiwu]

Review Summary

Well-targeted hardening of the OneHot operator across CPU and CUDA execution providers. The changes close real overflow and null-dereference vectors that could be triggered by adversarial model inputs. The implementation is correct, the test coverage is solid, and the code follows ORT conventions.

Positives:

• Manual overflow check for total output element count produces clear INVALID_ARGUMENT errors — good UX for operators accepting untrusted model data.
• SafeInt<int64_t> for prefix multiplication is proper belt-and-suspenders defense.
• CUDA int32 range validation before gsl::narrow_cast<int> is critical — silent truncation would produce wrong fast_divmod operands.
• Null output check correctly placed before output->Shape().Size() dereference on both CPU and CUDA paths.
• Comprehensive test coverage: overflow, negative depth, minimum depth, scalar rejection, and opset 9 backward compat.

Two minor suggestions below (non-blocking).

[tianleiwu]

One small build-hygiene nit below. The rest of the hardening looks good on the current head.