검색 상세

DP-FedVAE: Differentially Private Federated Variational Autoencoder for IoT Zero-Day Threat Detection

초록/요약

The significant growth in the number of Internet of Things devices has created serious security challenges, especially due to malware infections, botnet attacks, and zero-day threats. Conventional centralized machine-learning methods require collecting raw traffic data from distributed devices, which may introduce privacy risks, communication costs, and scalability limitations. Federated Learning （FL） offers a practical solution by enabling multiple clients to jointly train a shared model without directly exchanging raw local data. However, standard FL can still suffer from privacy leakage through exchanged model updates and can be exposed to malicious clients that attempt to manipulate the training process. To overcome these challenges, this thesis proposes DP-FedVAE, a privacy-aware federated anomaly-detection framework for zero-day threat detection in IoT environments. The proposed framework integrates a Variational Autoencoder （VAE） with FL and Differential Privacy （DP）. The VAE learns normal IoT traffic behavior through a probabilistic latent representation, making it suitable for detecting unseen or zero-day attacks using reconstruction error. DP is applied during federated aggregation through update clipping and Gaussian noise addition to reduce the exposure of exact client updates. However, the main experimental setting uses a noise multiplier of 𝜎 = 0.3, which prioritizes detection utility and robustness over strong formal privacy. Under the simplified RDP-based accounting assumptions used in this thesis, this setting gives an estimated privacy budget of 𝜖 = 344.85 at 𝛿 = 10−5, indicating weak formal privacy protection. Therefore, the primary role of the DP mechanism in this work is to study privacy-aware update perturbation and its robustness effect under adversarial federated learning, rather than to claim a strong formal privacy guarantee. A model-cancelling attack is also considered to assess the robustness of the proposed framework under adversarial federated-learning conditions. The experiments were performed on the N-BaIoT dataset using device-wise non-IID client par- titions. The proposed model was evaluated under four scenarios: clean, attack, DP-clean, and DP-attack, over 30 federated communication rounds. The experimental results indicate that FedVAE achieved 99.97% accuracy and 99.98% F1-score in the clean setting. However, under the model- cancelling attack, FedVAE’s performance dropped to 77.94% accuracy and 86.40% F1-score. In contrast, DP-FedVAE under attack achieved 81.59% accuracy and 88.90% F1-score on average over 30 federated communication rounds, improving accuracy by 3.65 percentage points and F1-score by 2.50 percentage points compared with the attacked FedVAE model. Its best observed performance during training reached 91.06% accuracy and 94.96% F1-score. These results demonstrate that DP-FedVAE provides a practical balance among anomaly detection performance, privacy preservation, and robustness against malicious clients. Therefore, the proposed framework is a promising approach for secure and privacy-aware IoT zero-day threat detection in distributed environments. Keywords: IoT, Federated Learning, Variational Autoencoder, Differential Privacy, zero-day threat detection, anomaly detection, IoT security.

more

목차

Chapter 1 Introduction 1
1.1 Background and Motivation 1
1.2 Research Gap, Objectives, and Contributions 5
1.3 Organization of the Thesis 6
Chapter 2 Literature Review and Research Gap 7
2.1 Overview 7
2.2 IoT Security and Anomaly Detection 8
2.3 Federated Learning, Privacy, and Robustness 11
2.4 Datasets and Research Gap 13
2.5 Chapter Summary 15
Chapter 3 System Model and Proposed Methodology 16
3.1 Overview 16
3.2 System Architecture and Data Assumptions 17
3.3 Proposed DP-FedVAE Learning Model 19
3.4 Federated Training and Privacy-Preserving Aggregation 21
3.5 Threat Model and Zero-Day Detection 26
3.6 Training Algorithm and Problem Formulation 27
3.7 Advantages over the Baseline Method 29
3.8 Chapter Summary 29
Chapter 4 Experimental Setup, Results, and Discussion 31
4.1 Experimental Overview and Dataset 31
4.2 Experimental Configuration 33
4.3 Average Performance over 30 Federated Rounds 37
4.4 Confusion Matrix Analysis 41
4.5 Discussion and Summary of Findings 48
4.6 Chapter Summary 49
Chapter 5 Conclusion and Future Work 51
5.1 Conclusion and Main Contributions 51
5.2 Limitations and Future Work 53
5.3 Final Remarks 55
Bibliography 56

more
반출 Meta View 목록