Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FedSecureFormer: A Fast, Federated and Secure Transformer Framework for Lightweight Intrusion Detection in Connected and Autonomous Vehicles

Devika S, Vishnu Hari, Pratik Narang, Tejasvi Alladi, F. Richard Yu

TL;DR

The paper tackles intrusion detection in Connected and Autonomous Vehicles under strict latency and privacy constraints. It introduces FedSecureFormer, a lightweight encoder-only transformer (1.7M parameters) trained in a federated, privacy-preserving setting, complemented by Hist-AttnGAN to probe unseen attacks and a DP mechanism to protect updates. Key contributions include state-of-the-art per-class recall on VeReMi Extension, real-time edge inference on Jetson Nano (3.7775 ms), and minimal federated accuracy loss under DP (≈4.04%), along with robust unseen-attack detection (88%). The work demonstrates practical viability for latency-sensitive ITS, enabling scalable, privacy-aware IDS deployment across vehicles and edge nodes.

Abstract

This works presents an encoder-only transformer built with minimum layers for intrusion detection in the domain of Connected and Autonomous Vehicles using Federated Learning.

FedSecureFormer: A Fast, Federated and Secure Transformer Framework for Lightweight Intrusion Detection in Connected and Autonomous Vehicles

TL;DR

The paper tackles intrusion detection in Connected and Autonomous Vehicles under strict latency and privacy constraints. It introduces FedSecureFormer, a lightweight encoder-only transformer (1.7M parameters) trained in a federated, privacy-preserving setting, complemented by Hist-AttnGAN to probe unseen attacks and a DP mechanism to protect updates. Key contributions include state-of-the-art per-class recall on VeReMi Extension, real-time edge inference on Jetson Nano (3.7775 ms), and minimal federated accuracy loss under DP (≈4.04%), along with robust unseen-attack detection (88%). The work demonstrates practical viability for latency-sensitive ITS, enabling scalable, privacy-aware IDS deployment across vehicles and edge nodes.

Abstract

This works presents an encoder-only transformer built with minimum layers for intrusion detection in the domain of Connected and Autonomous Vehicles using Federated Learning.
Paper Structure (26 sections, 25 equations, 8 figures, 9 tables, 2 algorithms)

This paper contains 26 sections, 25 equations, 8 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Literature Survey
  3. Preliminary Background
  4. CAV Scenario
  5. Dataset
  6. Transformers
  7. Federated Learning
  8. Proposed Intrusion Detection Framework
  9. Proposed FedSecureFormer Transformer
  10. Histogram-aware Attention GAN (Hist-AttnGAN)
  11. Federated Learning Framework with Differential Privacy
  12. Experimental Setup
  13. System Specification
  14. Dataset Specification
  15. Model Hyperparameters
  16. ...and 11 more sections

Figures (8)

  • Figure 1: CAV Scenario
  • Figure 2: FedSecureFormer architecture based on a 6-layer encoder-only transformer model, designed for efficient intrusion detection in CAVs, illustrated within a Federated Learning setup involving $n$ clients.
  • Figure 3: His-AttnGAN architecture utilized for generating unseen data.
  • Figure 4: FedSecureFormer achieves optimal performance across Accuracy, Precision, Recall, and F1-score plotted against the number of encoder layers. Red labels indicate the best values at six encoder layers.
  • Figure 5: FedSecureFormer achieves similar accuracy with both 2 and 4 attention heads, though 4 heads significantly increase training time.
  • ...and 3 more figures