💻 computer science

A Privacy-Aware and Communication-Efficient Federated Spam Detection Framework for Multi-Cloud Environments

This paper proposes FPSD-MCP, a novel federated learning framework that integrates secure multi-party computation and homomorphic encryption with an adaptive aggregation strategy to achieve a scalable, privacy-preserving, and communication-efficient spam detection solution for heterogeneous multi-cloud environments.

Original authors: Shanmuga Priya R, Yogesh Rajkumar R, Chellaswamy C

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shanmuga Priya R, Yogesh Rajkumar R, Chellaswamy C

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Secret Recipe" Dilemma

Imagine you are trying to teach a robot to spot spam emails. To do this well, the robot needs to see millions of emails from different people. However, there's a catch: nobody wants to show their private emails to anyone else. It's like trying to teach a chef to cook a perfect dish, but every chef in the world refuses to let the others see their secret family recipes.

Traditionally, to train a good spam detector, you would have to gather all those private emails into one giant central kitchen (a central server). But this is dangerous because if the kitchen gets robbed (hacked), everyone's private data is stolen. Also, moving all those emails takes a lot of time and bandwidth (like trying to ship a library of books across the ocean).

The Solution: The "Federated" Approach

The authors propose a new way called Federated Learning. Instead of bringing the recipes to a central kitchen, they send the chef (the AI model) to each person's house.

Local Training: The chef goes to Client A's house, learns from their emails, and writes down only the new cooking tips (mathematical updates) on a notepad. The actual emails stay locked in Client A's safe.
The Return Trip: The chef returns to the central hub with just the notepad, not the emails.
Aggregation: The central hub collects notepads from many clients, combines the tips, and creates a "Master Chef" who is smarter than any single chef.

The New Twist: The "Multi-Cloud" Challenge

The paper focuses on a specific, tricky version of this called Multi-Cloud. Imagine the clients aren't just neighbors; they are different companies using different cloud providers (like AWS, Google Cloud, Azure). These clouds speak different languages and have different rules.

The authors noticed that standard "Federated Learning" has two big problems in this messy environment:

Privacy Leaks: Even if you only send the "notepads" (model updates), a clever hacker can sometimes reverse-engineer the original emails from those notes.
Communication Clogs: Sending encrypted notes takes up a lot of space and time, slowing everything down.

The Proposed Framework: FPSD-MCP

The authors built a new system called FPSD-MCP. Think of it as a highly secure, efficient courier service for these "notepads." It uses three main tools to solve the problems:

1. The "Magic Envelope" (Homomorphic Encryption)

Imagine you put your notepad inside a magical, unbreakable glass box. You can shake the box, weigh it, or even do math on it while it is still inside the box, without ever opening it.

In the paper: This is Homomorphic Encryption (HE). It allows the central server to do calculations on the encrypted updates without ever seeing the raw data. The server is like a blindfolded accountant who can add up numbers without seeing what the numbers are.

2. The "Secret Split" (Secure Multi-Party Computation)

Imagine you don't send the whole notepad to one person. Instead, you tear it into 10 pieces and give one piece to 10 different people. No single person can read the note unless they all get together and combine their pieces.

In the paper: This is Secure Multi-Party Computation (MPC). It ensures that even if the central server is a bit shady, it can't figure out what any single client contributed unless it colludes with many others.

3. The "Smart Weighting" System

In a normal system, every client gets an equal vote, or the vote is based on how many emails they have. But what if a client has bad internet (high cost) or their data is very noisy?

In the paper: The authors created a smart weighting strategy. The system acts like a wise judge. It looks at:
- How much data the client has.
- How expensive it is to send their data (communication cost).
- How much "noise" (privacy protection) was added to their data.
- The Result: It gives more weight to clients who are efficient and have good data, and less weight to those who are slow or very noisy. This makes the final "Master Chef" smarter and faster.

What Did They Find? (The Results)

The researchers tested this system using two famous collections of emails: the Enron dataset (real corporate emails) and SpamAssassin (a standard spam collection). They compared their new system against older methods like standard Federated Averaging (FedAvg), LSTM, and BERT.

Here is the "scorecard" in plain English:

Better Accuracy: Their system caught more spam and let fewer spam emails through compared to the old methods. On the Enron dataset, they improved accuracy by about 4.5% and on SpamAssassin by 11.7%.
Faster Convergence: The system reached its "smartest" point in fewer rounds of training (fewer trips back and forth between clients and the server).
Less Traffic: Even though they were using heavy encryption (which usually slows things down), their "smart weighting" system actually used less bandwidth than some other secure methods.
Robustness: Even when they simulated bad actors trying to poison the data (sending fake notes to trick the system), FPSD-MCP held its ground much better than the other models.

The Bottom Line

The paper claims that FPSD-MCP is a way to build a super-smart spam filter that respects privacy. It allows different companies to work together to fight spam without ever sharing their private emails, while keeping the process fast and secure enough to work in a real-world, multi-cloud environment.

Key Takeaway: You can have your cake (high accuracy) and eat it too (strong privacy), without getting a stomach ache (slow performance), by using this new "smart, encrypted, and weighted" approach.

Technical Summary: A Privacy-Aware and Communication-Efficient Federated Spam Detection Framework for Multi-Cloud Environments (FPSD-MCP)

Problem Statement
The rapid proliferation of spam traffic in distributed multi-cloud environments presents significant challenges regarding scalability, communication efficiency, and data privacy. Traditional centralized spam detection approaches are increasingly unsuitable due to strict data-sharing constraints between organizations and cloud providers, as well as the high communication overhead associated with aggregating raw data. While Federated Learning (FL) offers a decentralized alternative by training models on local data without sharing raw inputs, standard FL approaches face critical limitations in multi-cloud settings. These include vulnerability to privacy leaks through gradient inversion attacks, degradation of model performance under heterogeneous (non-IID) data distributions, and the substantial computational and communication overhead introduced when integrating privacy-preserving mechanisms like Secure Multi-Party Computation (MPC) and Homomorphic Encryption (HE). Existing frameworks often treat privacy and communication efficiency as separate concerns, failing to provide a unified solution that balances model accuracy, privacy guarantees, and system overhead.

Methodology
The authors propose FPSD-MCP (Federated Privacy-Preserving Spam Detection for Multi-Cloud Platforms), a novel framework that integrates Federated Learning with MPC and HE to enable secure, scalable spam detection across heterogeneous cloud clusters.

Framework Architecture: The system operates within a multi-cloud cluster where each cloud provider acts as a client node with independent data and infrastructure. A Central Secure Router (CSR) coordinates the aggregation of model updates. The framework employs a semi-trusted assumption for the CSR, ensuring that even the coordinator cannot access plaintext model updates due to cryptographic protections.
Privacy Mechanisms:
- Differential Privacy (DP): Local gradients are clipped and perturbed with Gaussian noise to prevent information leakage.
- Homomorphic Encryption (HE): Privacy-preserved gradients are encrypted (using the Microsoft SEAL library) to allow arithmetic operations on ciphertexts without decryption.
- Secure Multi-Party Computation (MPC): Encrypted updates are processed using MPC protocols (via MP-SPDZ) to perform secure aggregation via secret sharing, ensuring that individual client contributions remain confidential during the aggregation process.
Adaptive Aggregation Strategy: A core component of FPSD-MCP is a privacy-aware and communication-efficient aggregation mechanism. Unlike standard FedAvg, which uses uniform weighting, FPSD-MCP formulates federated training as a multi-objective optimization problem. It adaptively weights client updates ( $w_i$ $w_{i}$ ) based on three factors:
1. Local Data Distribution: The size of the client's dataset ( $|D_i|$ ).
2. Communication Cost: The bandwidth and latency constraints ( $C_i$ ), which are higher under HE due to ciphertext expansion.
3. Privacy-Induced Noise: The variance of the differential privacy noise ( $\sigma_i^2$ ).
  This strategy down-weights clients with high communication costs or high privacy noise, optimizing the trade-off between model convergence, accuracy, and system efficiency.
Model Implementation: The spam detection model utilizes an LSTM-based classifier optimized for sequential email representation. The framework was tested using the Enron and SpamAssassin datasets, partitioned into non-IID client subsets to simulate realistic multi-cloud conditions.

Key Contributions
The paper outlines four primary contributions:

Novel Aggregation Strategy: A privacy-aware and communication-efficient weighting mechanism that adaptively adjusts client contributions based on data distribution, communication costs, and privacy noise, improving robustness under non-IID conditions.
Multi-Objective Optimization: A formulation that jointly minimizes classification loss, communication overhead, and privacy costs within a unified framework.
Integrated Privacy Framework: The combination of FL, MPC, and HE to ensure end-to-end privacy preservation and secure collaboration across heterogeneous cloud providers.
Comprehensive Evaluation: Extensive experimental validation demonstrating improved accuracy, faster convergence, and reduced communication overhead compared to state-of-the-art aggregation methods (FedAvg, FedProx, FedNova).

Experimental Results
The FPSD-MCP framework was evaluated on the Enron and SpamAssassin datasets against baseline models (Logistic Regression, LSTM, BERT) and standard aggregation strategies.

Performance Gains: FPSD-MCP significantly outperformed baseline models. On the Enron dataset, it achieved gains of 4.5% in precision, 3.6% in recall, and 4.1% in F1-score compared to existing methods. On the SpamAssassin dataset, improvements were even more pronounced, with gains of 11.7% in precision, 7.1% in recall, and 9.4% in F1-score.
Convergence and Efficiency: The proposed framework required fewer communication rounds to converge (28 rounds for Enron, 26 for SpamAssassin) compared to FedProx and FedNova, while maintaining lower communication overhead (16.3 MB for Enron, 14.1 MB for SpamAssassin).
Robustness:
- Non-IID Data: Under severe data distribution skew, FPSD-MCP maintained an F1-score of 0.90, significantly outperforming Logistic Regression (0.71) and BERT (0.84).
- Adversarial Attacks: In the presence of 30% poisoned data, FPSD-MCP showed high resilience, with an F1-score drop of only 0.06 (from 0.94 to 0.88), whereas Logistic Regression suffered a degradation of over 40%.
- Scalability: The framework maintained stable performance (96.8% accuracy) as the number of clients increased to 100, outperforming BERT and LSTM in large-scale settings.
Privacy-Efficiency Trade-off: The ablation study confirmed that while integrating MPC and HE introduces a minor accuracy drop (less than 3%) and increased computational cost compared to plaintext FL, the framework successfully balances these costs with strong security guarantees.

Significance and Claims
The authors claim that FPSD-MCP provides a scalable and secure solution for distributed spam detection, effectively addressing the limitations of current FL approaches in multi-cloud environments. The significance of the work lies in its unified design, which treats privacy and communication efficiency not as independent constraints but as jointly optimized objectives. By introducing an adaptive aggregation strategy that accounts for system-level trade-offs, the framework enables reliable model training across heterogeneous, non-IID data distributions without compromising data locality. The results suggest that FPSD-MCP is a viable candidate for real-world deployment in privacy-sensitive, multi-cloud infrastructures where balancing detection accuracy with strict privacy guarantees and communication constraints is critical. Future work is identified as focusing on further improving communication efficiency and optimizing resource allocation.