bastionlab

Mithril Security – BastionLab

Website | Documentation | Discord | Blog | LinkedIn | Twitter

Welcome to BastionLab!

Where data owners and data scientists can securely collaborate without exposing data - opening the way to projects that were too risky to consider.

⚙️ What is BastionLab?

BastionLab is a simple privacy framework for data science collaboration, covering data exploration and AI training.

It acts like an access control solution, for data owners to protect the privacy of their datasets, and stands as a guard, to enforce that only privacy-friendly operations are allowed on the data and anonymized outputs are shown to the data scientist.

• Data owners can let external or internal data scientists explore and extract values from their datasets, according to a strict privacy policy they'll define in BastionLab.
• Data scientists can remotely run queries on data frames and train their models without seeing the original data or intermediary results.

BastionLab is an open-source project. Our solution is coded in Rust , uses Polars , a pandas-like library for data exploration, and Torch , a popular library for AI training. We also have an option to set-up confidential computing , a hardware-based technology that ensures no one but the processor of the machine can see the data or the model.

Quick tour

You can go try out our Quick tour in the documentation to discover BastionLab with a hands-on example using the famous Titanic dataset.

But here’s a taste of what using BastionLab could look like

Data exploration

Data owner's side

…

Data scientist's side

…

AI training

Data owner's side

…

️ Key features

• Access control: data owners can define an interactive privacy policy that will filter the data scientist queries. They do not have to open unrestricted access to their datasets anymore.
• Limited expressivity: BastionLab limits the type of operations that can be executed by the data scientists to avoid arbitrary code execution.
• Transparent remote access: the data scientists never access the dataset directly. They only manipulate a local object that contains metadata to interact with a remotely hosted dataset. Calls can always be seen by data owners.

Getting help

• Go to our Discord #support channel
• Report bugs by opening an issue on our BastionLab Github
• Book a meeting with us

Disclaimer

BastionLab is still in development. Do not use it yet in a production workload. We will audit our solution in the future to attest that it enforces the security standards of the market.

License

BastionLab is licensed under the Apache License, Version 2.0.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

See the License for the specific language governing permissions and limitations under the License.