Progresses until now ( Mar 8, 2021):

1. Updated the project website.

2. Narrowed down the goal and Settled down the approach of the project.

3. Did research on the topic, read more papers and documents, and find the tools of implementation.

Why we made some change on the initial proposal:

Our current progress is on schedule, according to the initial project proposal, settling down the specific goal and the research approach. However, after digging into the problem and reading more papers, we made some changes to the proposal. First, we narrowed down the topic from “Anomaly Detection System” to “Classification” because anomaly detection as a topic is too complicated and needs more time and effort to fulfil. Meanwhile, classification is the foundation of anomaly detection, and they share the same algorithm, Federated Learning, to refine the model. After studying the improvement of classification, we found the refinement will help build the anomaly detection system. So the core of the whole project has not changed, and we just find a better angle to address the problem. Based on the more papers, we also found some papers already have had similar research, so we changed the “what has been done and ongoing projects” section. For the “approach” section, we made some changes as well, because, as more search has been done, we found more tools, approaches on this topic. However, the core idea is not changed, which is applying a decentralized approach to the Federated Learning on classification system.

What's the problem, and why is it important:

As we are trying to classify IoT devices and use learning to improve the model, we find that federated learning uses a centralized way to aggregate all the models from clients. An idea comes to us that why not try a decentralized way. In this project, we will explore how p2p network effects the efficiency of federated learning in the aggregation process.

Federated learning is widely used in mobile data processes, for the advantage of high security and outstanding performance on distributed systems. In the meantime, people put great efforts to improve it, which mainly focuses on overcoming statistical challenges, improving security as well as personalizing the model. We try to improve it in a different way, using a p2p network to aggregate the models rather than centrally gather models to the cloud.

What has been done, why are they not sufficient, and our other previous ongoing projects:

As mentioned before, most work to improve federated learning is about the statistical challenges, security and personalization. They have their own pros and cons, but the network framework may be a not-fully explored area for federated learning. Several papers are in Medical Imaging [3], healthcare informatics, generic area, etc. Some attempts have been made in IoT [4][5]. However, one focuses on certain situations like industrial IoT devices [5], and another achieves decentralization by exchange model with other nodes of one-hop distance.[5]. Our project focuses on the overall performance of decentralized federated learning in classification of IoT data flow. In addition, we will try several graphs of decentralization, including chain, which is similar to [5], tree, etc.

Approach, and why can it do better or differently:

The approaches are designing experiments, simulation, comparison among different classification models, and explanation. Moreover, the approach of implementation is a decentralized way to aggregate the classification model. Our project focused on a small office and home environment, which is different from IoT industries with large numbers of devices in one spot. It is hard to practice in reality, so we simulate this environment in a local computer, and deploy each spot with a node structure which has data, and one own model. These nodes will deliver their models to a central one or exchange models with pals. After rounds of iteration, we will compare the efficiency of centralized federated learning and decentralized one. In this way, the comparison of networks will be objective and rigorous. The combination of experiments, simulation, explanation, and comparison will highlight our approach to improve classification of IoT devices.

Expected deliverables:

The final deliverable product will be a report with charts and analysis. If necessary, we will submit our code, and trained models.

Biweekly time schedule:

TMar 8: finish and submit the project update (updated project proposal).

Mar 22: implement the codes and do the experiments.

Apr 5: finish the codes and show the demo.

Apr 12: submit the project report.

References:

[1] E. Lear, R. Droms, and D. Romascanu, “Manufacturer usage description specification,”

https://tools.ietf.org/html/draft-ietf-opsawg-mud-25

[2] S. Marchal, M. Miettinen, T. D. Nguyen, A. Sadeghi, and N. Asokan, “Audi: Towards autonomous iot device-type identification using periodic communication,” IEEE Journal on Selected Areas in Communications , pp. 1–1, 2019.

https://ieeexplore.ieee.org/abstract/document/8664655?casa_token=B9lJJXxPfW0AAAAA:icDR_f6BM2TPj_-xunh54bVrY-Q-nsXbWv-QrKgC3SWSjUfHzpyib1Cv2ZSirKzDVG2rfnCj

[3] Roy AG, Siddiqui S, Pölsterl S, Navab N, Wachinger C. Braintorrent: A peer-to-peer environment for decentralized federated learning. arXiv preprint arXiv:1905.06731. 2019 May 16. https://arxiv.org/abs/1905.06731

[4] S. Savazzi, M. Nicoli and V. Rampa, "Federated Learning With Cooperating Devices: A Consensus Approach for Massive IoT Networks," in IEEE Internet of Things Journal, vol. 7, no. 5, pp. 4641-4654, May 2020, doi: 10.1109/JIOT.2020.2964162.

https://ieeexplore.ieee.org/abstract/document/8950073?casa_token=qeZPtFqlOGwAAAAA:RLgJSVnc4k08NjVa9lTk3hbkiHqyFkjkWoApdOOGcv_guRR1VEolck582eGEQAPo8yi1F4sP

[5] Y. Lu, X. Huang, Y. Dai, S. Maharjan and Y. Zhang, "Blockchain and Federated Learning for Privacy-Preserved Data Sharing in Industrial IoT," in IEEE Transactions on Industrial Informatics, vol. 16, no. 6, pp. 4177-4186, June 2020, doi: 10.1109/TII.2019.2942190.

https://ieeexplore.ieee.org/abstract/document/8843900?casa_token=k9_2LXJ1jwAAAAA:f409ViEylxmvcVhMXaLV9LeTa8o78peintnmqRqd892S-FXfOlHHPXbca8j8PISTE06VorSx