FHE Based Mail Spam Detection

Fully homomorphic encryption based mail spam detection. Essentially, a client encrypts an email with its private key and sends it (encrypted) over to a server. The server, without decrypting the email, checks if it classifies as spam and returns an encrypted response to the client. Only the client with its private key can decrypt the response.

Spam classification at that stage just means checking if certain spam strings are part of the email message. If that is the case, we report spam. Otherwise, the message is good and it is not spam.

Works via Google FHE: https://github.com/google/fully-homomorphic-encryption. Google FHE is based on the TFHE library: https://tfhe.github.io/tfhe/.

I got the idea from Vitalik Buterin's post on FHE: https://vitalik.ca/general/2020/07/20/homomorphic.html

This example is a topic of a playlist on my YouTube channel: https://www.youtube.com/playlist?list=PLNEYVOFMCyfMGBz31YjrYKtuYXtceSCeg

C++ Code Generation

As Google's FHE library (and supposedly TFHE) has a limitation that doesn't allow dynamic arrays, we need to provide the spam strings at compile time. We do that by generating C++ code via Python. We have a number of C++ code templates that we just replace code in based on configuration and spam input. Then we pass the generated code to Google's FHE library and it compiles it to FHE code. That way we achieve "dynamic" array support.

The resulting pipeline is shown below:

Spam Strings

Spam strings are strings that we look for in emails. If an email contains one of them, the email is classified as spam. The spam.txt file contains one entry per line. The new line character is \n.

If a new line must be present, quoting can be used via the " character.

"new
line"

If the " character needs to be part of the spam string, it can be escaped via the \ character:

"contains \""

Usage

1. Install and compile Google's FHE library by following its README file.
2. The is_mail_spam directory contains a working example that is meant to be put in Google's FHE library transpiler/examples directory. Just copy or symlink the is_mail_spam directory.
3. Edit the spam.txt file so that it includes strings that are classified as spam.
4. Execute the generate_spam_detection_code.py Python script to generate C++ code.
5. Execute the unit test to make sure all is good:

bazelisk run //transpiler/examples/is_mail_spam:is_mail_spam_test 

6. Execute the testbench too. It simulates a client-server interaction:

bazelisk run //transpiler/examples/is_mail_spam:is_mail_spam_interpreted_tfhe_testbench "evil.url3"
...
[Client] Mail encryption done

                                        [Server] Computing
                                        [Server] Computation done
[Client] Decrypted result: 1
[Client] Mail is spam!

TODO

• Implement and actual client and a server. Use HTTP with JSON-RPC.