Deployment

Deployment process

After developing the model or converting the plaintext model into an FHE model, deploy it using Aegis's client and server features. In a typical Aegis deployment:

The server holds the artifacts generated by Aegis after compilation, specifically the server.zip file, which includes the program specification and FHE binary files (*.so).
The client holds the artifacts generated by Aegis after compilation, specifically the client.zip file, which includes the program specification.

The complete process is shown in the figure below:

deploy

Step-by-Step Deployment Guide

Follow these steps to develop, deploy, and use your privacy-preserving model.

Develop and compile the model (developer-side)

You can use the techniques discussed in the previous chapters to write your own model or reuse pre-trained models. The following code uses a pre-trained model.
```
from primus.aegis.fheruntime import FHEInferenceSession as InferenceSession
session = InferenceSession("/path/to/add_model.onnx")
```
The InferenceSession(...) function compiles the ONNX model and creates a session object to associate the compilation results.
Save deployment files (developer-side)

After compiling the model, save the required files for service deployment.
```
server_zip, client_zip = session.save()
```
The server_zip contains the program_spec.json and *.so files. The client_zip contains the program_spec.json.

The program_spec.json includes security parameters for generating the keyset, model input/output information, and other global configuration details. The *.so file is the compiled model binary.

After compilation completes, distribute the server.zip artifacts to the server and client.zip to client devices for deployment.
Load server.zip（server-side）

The server creates a service object using the server.zip file provided by the developer.
```
server = Server()
server = erver.load(server_zip)
```
Create client object (client-side)

After receiving the client_zip file sent by the developer-side, create a client object.
```
client = Client(client_zip)
```
Generate the keyset (client-side)

After completing the client object creation, perform key generation. The FHE CKKS scheme will generate the Private key, Galois keys, relinearization keys, and public key.
```
client.keygen()
```
Serialize the public key, Galois keys, and relinearization keys (client-side)
```
eva_keys = "/tmp/eva_keys.bin"
client.save_eva_keys(eva_keys)
```
Send the serialized public key, Galois keys, and relinearization keys to the server (client-side)

Encrypt the input data (client-side)

ser_enc_val1, ser_enc_val2 = client.encrypt([12.5, 76.3])

Send the serialized encrypted data to the server (client-side)
Deserialize the public key, Galois keys, and relinearization keys (server-side)
```
server.load_eva_keys(eva_keys)
```
Perform the computation and serialize the computation results (server-side)
```
ser_enc_output = server.run([enc_val1, enc_val2])
```
Send the serialized computation results to the client (server-side)
Deserialize the computation results and decrypt the computation results (client-side)
```
pt_output = client.decrypt(ser_enc_output)
```

Deployment process​

Step-by-Step Deployment Guide​

Deployment process

Step-by-Step Deployment Guide