dneprDroid/tfsecured

Small library for TensorFlow proto model's (*.pb) encryption/decryption

TFSecured

Small library for tensorflow proto models (*.pb) encryption/decryption.

AES

You may use random string with random length like a key, then library calculates sha256 hash of it and uses as internal key with size 256 bits.

Usage

Copy sources from TFSecured dir into your project.

C++ usage (see TFPredictor.mm):

    #include <GraphDefDecryptor.hpp>

    ........


    tensorflow::GraphDef graph;
    // Decryption: 
    const std::string key = "JHEW8F7FE6F8E76W8F687WE6F8W8EF5";
    auto status = tfsecured::GraphDefDecryptAES(path,         // path to *.pb file (frozen graph)
                                                &graph,
                                                key);         // your key
    if (!status.ok()) {
        std::cout << status.error_message() << std::endl;
        return;
    }
    
    // Create session :
    std::unique_ptr<Session> session(NewSession(options));
    status = session->Create(graph);
    
    // Run session ....

Encrypt model via python script (see encrypt_model.py):

$ python encrypt_model.py <INPUT_PB_MODEL>  \
                          <OUTPUT_PB_MODEL> \  
                          <KEY>                # optional, generated randomly by script 

Dependencies

• TensorFlow

• OpenSSL

iOS Demo (Digit recognizer)

Prepare repository (install Tensorflow via pods)

$ cd iosDemo
$ pod install 

And open it in Xcode.