This document discusses the TensorFlow security model. It describes how to safely deal with untrusted programs (models or model parameters), and input data. 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.

TensorFlow models (to use a term commonly used by machine learning practitioners) are expressed as programs that TensorFlow executes. TensorFlow programs are encoded as computation graphs. The model's parameters are often stored separately in checkpoints.

At runtime, TensorFlow executes the computation graph using the parameters provided. Note that the behavior of the computation graph may change depending on the parameters provided. TensorFlow itself is not a sandbox. When executing the computation graph, TensorFlow may read and write files, send and receive data over the network, and even spawn additional processes. All these tasks are performed with the permission of the TensorFlow process. Allowing for this flexibility makes for a powerful machine learning platform, but it has security implications.

The computation graph may also accept inputs. Those inputs are the data you supply to TensorFlow to train a model, or to use a model to run inference on the data.

TensorFlow models are programs, and need to be treated as such from a security perspective.

The TensorFlow library has a wide API which can be used in multiple scenarios. The security requirements are also different depending on the usage.

The API usage with the least security concerns is doing iterative exploration via the Python interpreter or small Python scripts. Here, only some parts of the API are exercised and eager execution is the default, meaning that each operation executes immediately. This mode is useful for testing, including fuzzing. For direct access to the C++ kernels, users of TensorFlow can directly call tf.raw_ops.xxx APIs. This gives control over all the parameters that would be sent to the kernel. Passing invalid combinations of parameters can allow insecure behavior (see definition of a vulnerability in a section below). However, these won’t always translate to actual vulnerabilities in TensorFlow. This would be similar to directly dereferencing a null pointer in a C++ program: not a vulnerability by itself but a coding error.

The next 2 modes of using the TensorFlow API have the most security implications. These relate to the actual building and use of machine learning models. Both during training and inference, the TensorFlow runtime will build and execute computation graphs from (usually Python) code written by a practitioner (using compilation techniques to turn eager code into graph mode). In both of these scenarios, a vulnerability can be exploited to cause significant damage, hence the goal of the security team is to eliminate these vulnerabilities or otherwise reduce their impact. This is essential, given that both training and inference can run on accelerators (e.g. GPU, TPU) or in a distributed manner.

Finally, the last mode of executing TensorFlow library code is as part of additional tooling. For example, TensorFlow provides a saved_model_cli tool which can be used to scan a SavedModel (the serialization format used by TensorFlow for models) and describe it. These tools are usually run by a single developer, on a single host, so the impact of a vulnerability in them is somewhat reduced.

As a general rule: Always execute untrusted models inside a sandbox (e.g., nsjail).

There are several ways in which a model could become untrusted. Obviously, if an untrusted party supplies TensorFlow kernels, arbitrary code may be executed. The same is true if the untrusted party provides Python code, such as the Python code that generates TensorFlow graphs.

Even if the untrusted party only supplies the serialized computation graph (in form of a GraphDef, SavedModel, or equivalent on-disk format), the set of computation primitives available to TensorFlow is powerful enough that you should assume that the TensorFlow process effectively executes arbitrary code. One common solution is to allow only a few safe Ops. While this is possible in theory, we still recommend you sandbox the execution.

It depends on the computation graph whether a user provided checkpoint is safe. It is easily possible to create computation graphs in which malicious checkpoints can trigger unsafe behavior. For example, consider a graph that contains a tf.cond operation depending on the value of a tf.Variable. One branch of the tf.cond is harmless, but the other is unsafe. Since the tf.Variable is stored in the checkpoint, whoever provides the checkpoint now has the ability to trigger unsafe behavior, even though the graph is not under their control.

In other words, graphs can contain vulnerabilities of their own. To allow users to provide checkpoints to a model you run on their behalf (e.g., in order to compare model quality for a fixed model architecture), you must carefully audit your model, and we recommend you run the TensorFlow process in a sandbox.

Similar considerations should apply if the model uses custom ops (C++ code written outside of the TensorFlow tree and loaded as plugins).

It is possible to write models that are secure in the sense that they can safely process untrusted inputs assuming there are no bugs. There are, however, two main reasons to not rely on this: First, it is easy to write models which must not be exposed to untrusted inputs, and second, there are bugs in any software system of sufficient complexity. Letting users control inputs could allow them to trigger bugs either in TensorFlow or in dependencies.

In general, it is good practice to isolate parts of any system which is exposed to untrusted (e.g., user-provided) inputs in a sandbox.

A useful analogy to how any TensorFlow graph is executed is any interpreted programming language, such as Python. While it is possible to write secure Python code which can be exposed to user supplied inputs (by, e.g., carefully quoting and sanitizing input strings, size-checking input blobs, etc.), it is very easy to write Python programs which are insecure. Even secure Python code could be rendered insecure by a bug in the Python interpreter, or in a bug in a Python library used (e.g., this one).

TensorFlow is a platform for distributed computing, and as such there is a TensorFlow server (tf.train.Server). The TensorFlow server is meant for internal communication only. It is not built for use in an untrusted network.

For performance reasons, the default TensorFlow server does not include any authorization protocol and sends messages unencrypted. It accepts connections from anywhere, and executes the graphs it is sent without performing any checks. Therefore, if you run a tf.train.Server in your network, anybody with access to the network can execute what you should consider arbitrary code with the privileges of the process running the tf.train.Server.

When running distributed TensorFlow, you must isolate the network in which the cluster lives. Cloud providers provide instructions for setting up isolated networks, which are sometimes branded as "virtual private cloud." Refer to the instructions for GCP and AWS) for details.

Note that tf.train.Server is different from the server created by tensorflow/serving (the default binary for which is called ModelServer). By default, ModelServer also has no built-in mechanism for authentication. Connecting it to an untrusted network allows anyone on this network to run the graphs known to the ModelServer. This means that an attacker may run graphs using untrusted inputs as described above, but they would not be able to execute arbitrary graphs. It is possible to safely expose a ModelServer directly to an untrusted network, but only if the graphs it is configured to use have been carefully audited to be safe.

Similar to best practices for other servers, we recommend running any ModelServer with appropriate privileges (i.e., using a separate user with reduced permissions). In the spirit of defense in depth, we recommend authenticating requests to any TensorFlow server connected to an untrusted network, as well as sandboxing the server to minimize the adverse effects of any breach.

It is possible to run multiple TensorFlow models in parallel. For example, ModelServer collates all computation graphs exposed to it (from multiple SavedModel) and executes them in parallel on available executors. A denial of service caused by one model could bring down the entire server, but we don't consider this as a high impact vulnerability, given that there exists solutions to prevent this from happening (e.g., rate limits, ACLs, monitors to restart broken servers).

However, it is a critical vulnerability if a model could be manipulated such that it would output parameters of another model (or itself!) or data that belongs to another model.

Models that also run on accelerators could be abused to do hardware damage or to leak data that exists on the accelerators from previous executions, if not cleared.

TensorFlow is a large and complex system. It also depends on a large set of third party libraries (e.g., numpy, libjpeg-turbo, PNG parsers, protobuf). It is possible that TensorFlow or its dependencies may contain vulnerabilities that would allow triggering unexpected or dangerous behavior with specially crafted inputs.

Given TensorFlow's flexibility, it is possible to specify computation graphs which exhibit unexpected or unwanted behavior. The fact that TensorFlow models can perform arbitrary computations means that they may read and write files, communicate via the network, produce deadlocks and infinite loops, or run out of memory. It is only when these behaviors are outside the specifications of the operations involved that such behavior is a vulnerability.

A FileWriter writing a file is not unexpected behavior and therefore is not a vulnerability in TensorFlow. A MatMul allowing arbitrary binary code execution is a vulnerability.

This is more subtle from a system perspective. For example, it is easy to cause a TensorFlow process to try to allocate more memory than available by specifying a computation graph containing an ill-considered tf.tile operation. TensorFlow should exit cleanly in this case (it would raise an exception in Python, or return an error Status in C++). However, if the surrounding system is not expecting the possibility, such behavior could be used in a denial of service attack (or worse). Because TensorFlow behaves correctly, this is not a vulnerability in TensorFlow (although it would be a vulnerability of this hypothetical system).

As a general rule, it is incorrect behavior for TensorFlow to access memory it does not own, or to terminate in an unclean way. Bugs in TensorFlow that lead to such behaviors constitute a vulnerability.

One of the most critical parts of any system is input handling. If malicious input can trigger side effects or incorrect behavior, this is a bug, and likely a vulnerability.

Note: Assertion failures used to be considered a vulnerability in TensorFlow. If an assertion failure only leads to program termination and no other exploits, we will no longer consider assertion failures (e.g., CHECK-fails) as vulnerabilities. However, if the assertion failure occurs only in debug mode (e.g., DCHECK) and in production-optimized mode the issue turns into other code weakness(e.g., heap overflow, etc.), then we will consider this to be a vulnerability. We recommend reporters to try to maximize the impact of the vulnerability report (see also the Google VRP rules and the Google OSS VRP rules).

Note: Although the iterative exploration of TF API via fuzzing tf.raw_ops.xxx symbols is the best way to uncover code weakness, please bear in mind that this is not a typical use case that has security implications. It is better to try to translate the vulnerability to something that can be exploited during training or inference of a model (i.e., build a model that when given a specific input would produce unwanted behavior). Alternatively, if the TensorFlow API is only used in ancillary tooling, consider the environment where the tool would run. For example, if saved_model_cli tool would crash on parsing a SavedModel that is not considered a vulnerability but a bug (since the user can use other ways to inspect the model if needed). However, it would be a vulnerability if passing a SavedModel to saved_model_cli would result in opening a new network connection, corrupting CPU state, or other forms of unwanted behavior.

Please use Google Bug Hunters reporting form to report security related issues.

Please use a descriptive title for your report.

In addition, please include the following information along with your report:

• Your name and affiliation (if any).
• A description of the technical details of the vulnerabilities. It is very important to let us know how we can reproduce your findings.
• A minimal example of the vulnerability.
• An explanation of who can exploit this vulnerability, and what they gain when doing so -- write an attack scenario. This will help us evaluate your report quickly, especially if the issue is complex.
• Whether this vulnerability is public or known to third parties. If it is, please provide details.

After the initial reply to your report, the security team will endeavor to keep you informed of the progress being made towards a fix and announcement. TensorFlow uses the following disclosure process:

• When a report is received, we confirm the issue and determine its severity. Please try to maximize impact in the report, going beyond just obtaining unwanted behavior in a fuzzer.
• If we know of specific third-party services or software based on TensorFlow that require mitigation before publication, those projects will be notified.
• An advisory is prepared (but not published) which details the problem and steps for mitigation.
• The vulnerability is fixed and potential workarounds are identified.
• We will publish a security advisory for all fixed vulnerabilities.

For each vulnerability, we try to ingress it as soon as possible, given the size of the team and the number of reports. Vulnerabilities will, in general, be batched to be fixed at the same time as a quarterly release.

Security advisories from 2018 to March 2023 are listed here. From TF 2.13 onwards, we have sunset this list and only use GitHub's Security Advisory format, to simplify the post-vulnerability-fix process. In both locations, we credit reporters for identifying security issues, although we keep your name confidential if you request it.