IEEE MDM 2023 Participation Report(Muhammad Ayat Hidayat)

IEEE MDM 2023 Participation Report

Outline of the Presentation

1. Oral Presentation

Title: AGC-DP: Differential Privacy with Adaptive Gaussian Clipping for Federated Learning

Content: We applied differential privacy for federated learning, which adds noise to the gradient of the local model before sending it to the central server for aggregation. We proposed AGC-DP, or Adaptive Gaussian Clipping, which consists of Adaptive Clipping, Gaussian Noise, and Sub-sampling. We also enhanced the privacy protection by using Privacy Loss Distribution (PLD). By utilizing all these components, we successfully achieved higher accuracy with high noise (High Privacy), especially in non-i.i.d datasets, compared to other differential privacy algorithms such as RDP, ZcDP, and DP-SGD. However, we still have a high total communication cost generated from the proposed method. In the future, we plan to implement compression techniques to reduce the communication cost of AGC-DP.

2. Demo Presentation

Title: Efficient and Secure: Privacy-preserving Federated Learning for Resource-Constrained Devices

Content: We applied compressive sensing to AGC-DP by compressing the local model's weight layer by layer before sending it to the central server. We perform the compression by multiplying the model's weight with the DCT matrix and then reconstructing the original weight on the central server using a reconstruction algorithm before updating the global model. By implementing this method, we can achieve higher accuracy with lower CPU time. This means our proposed method can be implemented on resource-constrained devices that have low computational capability and memory.

Experience

I had an amazing experience during my participation in the conference. Our journey began on July 3rd when we arrived in Singapore in the afternoon. The real excitement started on the second day, July 4th, when my professor and I attended the conference. Afterward, we embarked on a tour of some of Singapore's most iconic landmarks. We visited the Merlion statue and Marina Bay Sands, immersing ourselves in the vibrant atmosphere of the city.

As the day progressed, we found ourselves drawn to the Night Market, a bustling hub not far from the Merlion statue. The market offereda wide variety of local Singaporean, Malay, and Thai cuisines. It was a pleasure to savor the flavors and experience the diverse culinary traditions. For our Muslim friends, there were also options for halal food, ensuring everyone could enjoy the vibrant food scene.  

The following day was equally engaging as we demonstrated the application of security and privacy in federated learning. It was a gratifying experience to showcase our work and exchange knowledge with fellow researchers. To conclude the day, we had the opportunity to indulge in the flavors of Indian cuisine at the renowned Yantra Restaurant, enhancing our culinary adventure.  

On the fourth and final day, after the exhilarating oral presentations, we embarked on a boat ride along the river, soaking in the picturesque views of the city. Our journey led us to the breathtaking Gardens by the Bay, where we were captivated by the extraordinary flora housed within the domes. It was a truly enchanting experience, and we were in awe of the beauty of nature.  

During the conference, I had the opportunity to delve into new research topics related to federated learning. One fascinating area was its utilization in Analyzing spatiotemporal data, opening up new possibilities for understanding complex patterns. Additionally, exploring the implementation of federated learning in Low Earth Orbit satellites sparked my curiosity about the future of distributed machine learning in space. Overall, my time in Singapore for the conference was an unforgettable experience filled with enriching academic insights, delectable culinary adventures, and awe-inspiring cultural landmarks. It was a journey that broadened my horizons both professionally and personally, leaving me with cherished memories that will last a lifetime.