Workflowdrift-adaptive-bert
Drift-Adaptive BERT
A real-time BERT-based network intrusion detection system with adaptive feedback loops, achieving 99.97% accuracy and 5.54ms average latency on evolving cyber threats.
stack
- Python
- PyTorch
- Transformers
- Docker
- FastAPI
Architecture
- BERT encoder — fine-tuned on network traffic features for binary benign/malicious classification
- Adaptive feedback loop — analyst-flagged misclassifications drive lightweight fine-tuning of the classification head only; encoder preserved
- Experience replay buffer — new feedback mixed with random historical training data each cycle to prevent forgetting older attacks
- Validation gate — updated model must pass a held-out test set before promotion to production
- Real-time inference — FastAPI serving layer with batched inference
- Containerized microservices — separate Docker containers for inference, feedback collection, and model update orchestration
Results
- 99.97% accuracy, 99.90% recall on evolving threat datasets
- 5.54ms average latency, viable for real-time traffic inspection
- Outperforms static models on novel attack detection after drift
Action · Send message
Hiring for a remote AI engineering role?
Contact me