# Using TensorFlow Securely

This document discusses the TensorFlow security model. It describes the security
risks to consider when using models, checkpoints or input data for training or
serving. We also provide guidelines on what constitutes a vulnerability in
TensorFlow and how to report them.

This document applies to other repositories in the TensorFlow organization,
covering security practices for the entirety of the TensorFlow ecosystem.

    absl::Span<const int64_t> input_tensor_id,
    absl::Span<const tensorflow::DataType> input_dtypes,
    const std::function<BackwardFunction*()>& backward_function_getter,
    const std::function<void(BackwardFunction*)>& backward_function_deleter) {
  if (!ShouldRecord(input_tensor_id, input_dtypes)) {
    return;
  }
  std::vector<int64_t> ids;
  ids.reserve(input_tensor_id.size());

    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }
    return absl::OkStatus();
  }
  ~IdentityNGradientFunction() override {}
};

        "//tensorflow/core/grappler/clusters:single_machine",
        "//tensorflow/core/grappler/costs:graph_properties",
        "//tensorflow/core/grappler/inputs:trivial_test_graph_input_yielder",
        "//tensorflow/core/protobuf:error_codes_proto_impl_cc",
    ],

# Usage: Go to
# https://github.com/tensorflow/tensorflow/actions/workflows/release-branch-cherrypick.yml
# and click "Run Workflow." Leave "Use Workflow From" set to "master", then
# input the branch name and paste the cherry-pick commit and click Run. A PR
# will be created.

name: Release Branch Cherrypick
on:
  workflow_dispatch:
    inputs:

//     OK to not have the inputs properly set in `op`. See `input_tensors`
//     if you want shape inference to consider the input tensors of the
//     op for shape inference.
//   - The types need not be set in `input_shapes` as it is not used.
//   - The number of `input_tensors` should be the same as the number of items
//     in `input_shapes`.
//
// The results are returned in `output_shapes` and

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // Grad for Softmax Input
    TF_RETURN_IF_ERROR(BroadcastMul(
        ctx, grad_outputs[0], forward_outputs_[1],
        grad_inputs.subspan(0, 1)));  // upstream_grad * local softmax grad

    // Grad for labels is null

    Model model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, int input_index,
    float* expected_grad, int num_grad, bool use_function,
    double abs_error = 1e-2) {
  Status s;
  AbstractTensorHandlePtr numerical_grad;
  {
    AbstractTensorHandle* numerical_grad_raw;
    s = CalcNumericalGrad(ctx, model, inputs, input_index, use_function,
                          &numerical_grad_raw);

struct ForwardOperation {
 public:
  string op_name;
  std::vector<AbstractTensorHandle*> inputs;
  std::vector<AbstractTensorHandle*> outputs;
  std::vector<int64_t> skip_input_indices;
  AttrBuilder attrs;
};

using GradientFunctionFactory =
    std::function<GradientFunction*(const ForwardOperation& op)>;

// Map from op name to a `GradientFunctionFactory`.
class GradientRegistry {

  // existing and given constraints will be performed.
  virtual Status SetDeviceName(const char* name) = 0;

  virtual Status AddInput(AbstractTensorHandle* input) = 0;
  virtual Status AddInputList(
      absl::Span<AbstractTensorHandle* const> inputs) = 0;
  virtual Status Execute(absl::Span<AbstractTensorHandle*> retvals,
                         int* num_retvals) = 0;