Introduction
Intrusion detection is core to enterprise security, but for BT as a security cloud provider, it faces a structural limitation: each customer site's threat detection runs in isolation. Attack patterns observed on one customer's network can't inform detection for another, and centralizing that data to fix the problem creates serious privacy and compliance risk.
To solve this, BT and PryvX built a Federated Intrusion Detection System (F-IDS) with homomorphic encrypted inference — letting BT's customer sites collaboratively train a shared threat-detection model without any raw network data ever leaving their premises, and without BT itself seeing that data even during inference.
The Approach
The system combines two techniques:
Federated learning (FL): Instead of pooling raw traffic logs, each customer site trains a local model on its own data. Only the trained model — not the underlying data — is shared with BT for aggregation into a global detection model.
Homomorphic encryption (HE): Even when BT hosts the model for inference, incoming traffic data stays encrypted throughout the entire computation. BT can run detection without ever seeing plaintext data, and only the customer holding the decryption key can read the result.
Together, these let BT offer a continuously improving, collective defense model while every customer retains full ownership and confidentiality of their data.
How It Works
Each customer site trains a Random Forest classifier locally, chosen for its robustness and how well it aggregates across independently trained models. These local models are sent to a central BT service, which combines their predictions through a federated ensemble (majority voting) — without retraining or ever accessing raw data.
For the encrypted path, each client generates a Paillier encryption key pair. Feature data is encrypted before it's sent, inference runs directly on the encrypted values, and only the client can decrypt the final prediction. Paillier's additive properties are a natural fit here, since ensemble inference is fundamentally about summing votes and scores across models.
Results
The system was validated against the UNSW-NB15 intrusion detection benchmark, split across simulated customer sites each seeing different attack types.
The federated ensemble model reached ~92–93% accuracy, outperforming any individual client's local model and matching a fully centralized (non-private) baseline.
The homomorphic inference pipeline matched plaintext accuracy almost exactly (93.24% vs. 93.30%) — confirming encryption doesn't meaningfully cost detection quality.
The trade-off is speed: encrypted inference runs roughly 20–30x slower than plaintext (tens of milliseconds vs. a few milliseconds per sample), with larger payloads due to ciphertext size. This overhead is acceptable for environments where confidentiality matters more than raw latency.
Why It's Secure
Federated learning alone still carries risks — model inversion, membership inference, and poisoning attacks. Layering in Paillier encryption closes the biggest gap: since the server never has access to decrypted inputs or intermediate results, it can't leak data even if compromised. Combined with federated training, this extends privacy guarantees all the way from training through inference, and aligns with frameworks like GDPR, the UK Data Protection Act, and the NIST Privacy Framework.
What's Next
The next phase focuses on integrating the system into real BT network environments and further optimizing the encrypted inference pipeline to close the latency gap with plaintext.
Conclusion
This project shows that privacy and collective defense don't have to be a trade-off. By combining federated learning with homomorphic encryption, BT can give customers a shared, constantly improving intrusion detection model — at 93%+ accuracy — without any customer ever having to expose raw network data to BT or to each other.
