【Machine Lewarning】 Paper Presentation
- 1. Distributed Anomaly Detection using Autoencoder Neural Networks in WSN for IoT SPEAKER: Shun-Yu, Ko ID:107598068
- 2. Outline 1. Introduction 2. Anomaly detection 3. Autoencoder neural network 4. System architecture and two-part algorithm 5. Performance evaluation 6. Results 7. Conclusion 8. Reference
- 3. Introduction
- 4. Introduction
- 5. Introduction 1. In this paper, we propose to use autoencoder neural networks for anomaly detection in WSN. 2. It has the minimal communication and computation requirements and the distributed advantage. 3. This work is the first to introduce Autoencoder into WSN. 5
- 8. The hyper parameters • Neural Network=input+Weight+Bias+Activation function 8
- 10. WSN Computational complexity • MAX Dimension : M ( Ex : 720 ) • Computation time : O(M2) According to the paper : Scalable hypergrid k-Nnbased online anomaly detection in wireless sensor networks , IEEE Transactions on Parallel and Distributed Systems, vol. 24, no. 8, pp. 1661–1670, 2013 ,it cost O(2M-1) Computation time ! 10
- 11. PERFORMANCE EVALUATION WSN Testbed and Dataset 1. 8 sensor nodes that monitor temperature and relative humidity. 2. 720 daily readings from a single sensor 3. Run for 4 months. 11
- 12. ROC Curve(receiver operating characteristic curve) 12 有高血壓 無高血壓
- 13. ROC Curve • Accuracy = (TP + TN) / (P + N) 13
- 14. Area under the Curve of ROC (AUC ROC)） 14
- 15. Results- Varying anomaly magnitude: 1. Vary the magnitude v of spikes and bursts . 2. Fixing the frequency of anomalies at K = 100 per day. 3. AUC > 0.8 most of the time, which indicates a good classifier. 15
- 16. Results- Varying anomaly frequency: 16
- 17. CONCLUSION 1. This paper presents the first effort of introducing autoencoder neural networks into WSN to perform anomaly detection. 2. Communication overhead is minimized and computational load is allocated to the most suitable entities. 3. Single hidden layer of neurons and the corresponding computational complexity is only polynomial O(M2) in order to suit resource-limited sensors. 4. Unsupervised learning is able to adapt to unforeseeable and new changes in a non-stable environment. 17
- 18. Reference • T. Luo and S. G. Nagarajan, "Distributed Anomaly Detection Using Autoencoder Neural Networks in WSN for IoT," 2018 IEEE International Conference on Communications (ICC), Kansas City, MO, 2018, pp. 1-6. doi: 10.1109/ICC.2018.8422402 18
Editor's Notes
- #5: The task of anomaly detection could be undertaken by a central IoT entity such as “back-end” as often referred to, such a scheme tends to cause highly inefficient resource utilization. Because of the large amount of data that needs to be transmitted from sensors to the central entity, which entails substantial channel interference and energy consumption. 1.異常檢測可以利用集中式的方式進行，但往往導致資源利用非常沒有效率。 2.因為Sensor需要持續傳輸data到sink點，需要大量的通道干擾和能源的消耗。 而且事實上只有很小一部分的資料是異常的