This is an automated email from the ASF dual-hosted git repository. jxie pushed a commit to branch master in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/master by this push: new eb72d72 Added a security best practices doc (#8744) eb72d72 is described below commit eb72d72707eae391d8d4a104e22be97950eee269 Author: Madan Jampani <mad...@users.noreply.github.com> AuthorDate: Tue Nov 21 16:52:58 2017 -0800 Added a security best practices doc (#8744) * Added a security best practices doc * Minor edit --- docs/faq/index.md | 3 +++ docs/how_to/security.md | 24 ++++++++++++++++++++++++ 2 files changed, 27 insertions(+) diff --git a/docs/faq/index.md b/docs/faq/index.md index 1bfaea4..e29bda0 100644 --- a/docs/faq/index.md +++ b/docs/faq/index.md @@ -40,6 +40,9 @@ and full working examples, visit the [tutorials section](../tutorials/index.md). * [How to convert MXNet models to Apple CoreML format?](https://github.com/apache/incubator-mxnet/tree/master/tools/coreml) +## Security +* [How to run MXNet securely?](http://mxnet.io/how_to/security.md) + ## Extend and Contribute to MXNet * [How do I join the MXNet development discussion?](http://mxnet.io/community/mxnet_channels.html) diff --git a/docs/how_to/security.md b/docs/how_to/security.md new file mode 100644 index 0000000..6f64a9e --- /dev/null +++ b/docs/how_to/security.md @@ -0,0 +1,24 @@ +# MXNet Security best practices + +MXNet framework has no built-in security protections. It assumes that the MXNet entities involved in model training and inferencing (hosting) are fully trusted. It also assumes that their communications cannot be eavesdropped or tampered with. MXNet consumers shall ensure that the above assumptions are met. + +In particular the following threat-vectors exist when training using MXNet: + +* When running distributed training using MXNet there is no built-in support for authenticating cluster nodes participating in the training job. +* Data exchange between cluster nodes happens is in plain-text. +* Using `kvstore.set_optimizer` one can use a custom optimizer to combine gradients. This optimizer code is sent to the server nodes as a pickle file. A server does not perform any further validation of the pickle file and simply executes the code trusting the sender (worker). +* Since there is no authentication between nodes, a malicious actor running on the same network can launch a Denial of Service (DoS) attack by sending data that can overwhelm/crash a scheduler or other server nodes. + +It is highly recommended that the following best practices be followed when using MXNet: + +* Run MXNet with least privilege, i.e. not as root. +* Run MXNet training jobs inside a secure and isolated environment. If you are using a cloud provider like Amazon AWS, running your training job inside a [private VPC] (https://aws.amazon.com/vpc/) is a good way to accomplish this. Additionally, configure your network security settings so as to only allow connections that the cluster nodes require. +* Make sure no unauthorized actors have physical or remote access to the nodes participating in MXNet training. +* During training, one can configure MXNet to periodically save model checkpoints. To protect these model checkpoints from unauthorized access, make sure the checkpoints are written out to an encrypted storage volume, and have a provision to delete checkpoints that are no longer needed. +* When sharing trained models, or when receiving trained models from other parties, ensure that model artifacts are authenticated and integrity protected using cryptographic signatures, thus ensuring that the data received comes from trusted sources and has not been maliciously (or accidentally) modified in transit. +* By default, mx.random uses a static and fixed seed value. The random utilities in MXNet should therefore never be used to implement any type of security critical functionality where cryptographically secure pseudorandom number generation is required. + +# Deployment Considerations +The following are not MXNet framework specific threats but are applicable to Machine Learning models in general. + +* When deploying high-value, proprietary models for inference, care should be taken to prevent an adversary from stealing the model. The research paper [Stealing Machine Learning Models via Prediction APIs] (https://arxiv.org/pdf/1609.02943.pdf) outlines experiments performed to show how an attacker can use a prediction API to leak the ML model or construct a nearly identical replica. A simple way to thwart such an attack is to not expose the prediction probabilities to a high degree of [...] -- To stop receiving notification emails like this one, please contact ['"comm...@mxnet.apache.org" <comm...@mxnet.apache.org>'].