Advanced Spam Detection with Homomorphic Encryption Technology

SpamGuard AI provides secure and privacy-preserving spam classification using homomorphic encryption (HE).

• Privacy-preserving spam classification using encrypted vectors and weights
• Seamless integration of preprocessing, vectorization, encryption, and prediction
• Designed for secure environments where data confidentiality is critical
• Client-side secret key management with no server-side exposure
• Supports encrypted inference with logistic regression over FHE

1. Client-side preprocessing
   The user’s input email text is preprocessed and converted into a numerical feature vector.

2. Client-side encryption
   The input vector and model parameters (weights and intercept) are encrypted using the piheaan library.
   ➤ The secret key is kept solely on the client side, and the server has no access to it.

3. Encrypted data transmission
   The encrypted input text(vector) and model parameters are Base64-encoded and transmitted to the server.

4. Server-side encrypted prediction
   The server performs logistic regression prediction directly on the encrypted data.
   ➤ The server never accesses any plaintext data during this computation.(NO secret key)

5. Encrypted result transmission
   The encrypted prediction result (a score) is returned to the client in Base64-encoded form.

6. Client-side decryption & classification
   The client decodes and decrypts the score using the secret key and classifies the result as spam or not spam.

   💡 Why not use the sigmoid function?
   This model performs binary classification, where the classification decision depends only on whether the probability is above or below 0.5.
   Therefore, it's not necessary to explicitly compute the probability using the sigmoid function — we can directly use the raw linear output z (before applying sigmoid) to make the decision.

   In a homomorphic encryption (HE) setting, approximating the sigmoid function is computationally expensive and introduces loss of precision due to its non-linearity and division operations.

   ⚙️ Instead of applying an expensive sigmoid approximation, we simply interpret the encrypted result based on whether z > 0 (not spam) or z ≤ 0 (spam).

• piheaan: HE library for homomorphic operations
• Logistic Regression: Machine learning model for binary classification
• React + TypeScript: Client-side application
• Python (FastAPI): Server-side server for prediction / Client-side server for encryption & decryption

All computations on the server are done on encrypted data. The secret key never leaves the client, making this system suitable for environments where data privacy and zero-trust architectures are required.